Merge tag 'regulator-fix-v5.5-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-microblaze.git] / fs / ext4 / indirect.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/ext4/indirect.c
4  *
5  *  from
6  *
7  *  linux/fs/ext4/inode.c
8  *
9  * Copyright (C) 1992, 1993, 1994, 1995
10  * Remy Card (card@masi.ibp.fr)
11  * Laboratoire MASI - Institut Blaise Pascal
12  * Universite Pierre et Marie Curie (Paris VI)
13  *
14  *  from
15  *
16  *  linux/fs/minix/inode.c
17  *
18  *  Copyright (C) 1991, 1992  Linus Torvalds
19  *
20  *  Goal-directed block allocation by Stephen Tweedie
21  *      (sct@redhat.com), 1993, 1998
22  */
23
24 #include "ext4_jbd2.h"
25 #include "truncate.h"
26 #include <linux/dax.h>
27 #include <linux/uio.h>
28
29 #include <trace/events/ext4.h>
30
31 typedef struct {
32         __le32  *p;
33         __le32  key;
34         struct buffer_head *bh;
35 } Indirect;
36
37 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
38 {
39         p->key = *(p->p = v);
40         p->bh = bh;
41 }
42
43 /**
44  *      ext4_block_to_path - parse the block number into array of offsets
45  *      @inode: inode in question (we are only interested in its superblock)
46  *      @i_block: block number to be parsed
47  *      @offsets: array to store the offsets in
48  *      @boundary: set this non-zero if the referred-to block is likely to be
49  *             followed (on disk) by an indirect block.
50  *
51  *      To store the locations of file's data ext4 uses a data structure common
52  *      for UNIX filesystems - tree of pointers anchored in the inode, with
53  *      data blocks at leaves and indirect blocks in intermediate nodes.
54  *      This function translates the block number into path in that tree -
55  *      return value is the path length and @offsets[n] is the offset of
56  *      pointer to (n+1)th node in the nth one. If @block is out of range
57  *      (negative or too large) warning is printed and zero returned.
58  *
59  *      Note: function doesn't find node addresses, so no IO is needed. All
60  *      we need to know is the capacity of indirect blocks (taken from the
61  *      inode->i_sb).
62  */
63
64 /*
65  * Portability note: the last comparison (check that we fit into triple
66  * indirect block) is spelled differently, because otherwise on an
67  * architecture with 32-bit longs and 8Kb pages we might get into trouble
68  * if our filesystem had 8Kb blocks. We might use long long, but that would
69  * kill us on x86. Oh, well, at least the sign propagation does not matter -
70  * i_block would have to be negative in the very beginning, so we would not
71  * get there at all.
72  */
73
74 static int ext4_block_to_path(struct inode *inode,
75                               ext4_lblk_t i_block,
76                               ext4_lblk_t offsets[4], int *boundary)
77 {
78         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
79         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
80         const long direct_blocks = EXT4_NDIR_BLOCKS,
81                 indirect_blocks = ptrs,
82                 double_blocks = (1 << (ptrs_bits * 2));
83         int n = 0;
84         int final = 0;
85
86         if (i_block < direct_blocks) {
87                 offsets[n++] = i_block;
88                 final = direct_blocks;
89         } else if ((i_block -= direct_blocks) < indirect_blocks) {
90                 offsets[n++] = EXT4_IND_BLOCK;
91                 offsets[n++] = i_block;
92                 final = ptrs;
93         } else if ((i_block -= indirect_blocks) < double_blocks) {
94                 offsets[n++] = EXT4_DIND_BLOCK;
95                 offsets[n++] = i_block >> ptrs_bits;
96                 offsets[n++] = i_block & (ptrs - 1);
97                 final = ptrs;
98         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
99                 offsets[n++] = EXT4_TIND_BLOCK;
100                 offsets[n++] = i_block >> (ptrs_bits * 2);
101                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
102                 offsets[n++] = i_block & (ptrs - 1);
103                 final = ptrs;
104         } else {
105                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
106                              i_block + direct_blocks +
107                              indirect_blocks + double_blocks, inode->i_ino);
108         }
109         if (boundary)
110                 *boundary = final - 1 - (i_block & (ptrs - 1));
111         return n;
112 }
113
114 /**
115  *      ext4_get_branch - read the chain of indirect blocks leading to data
116  *      @inode: inode in question
117  *      @depth: depth of the chain (1 - direct pointer, etc.)
118  *      @offsets: offsets of pointers in inode/indirect blocks
119  *      @chain: place to store the result
120  *      @err: here we store the error value
121  *
122  *      Function fills the array of triples <key, p, bh> and returns %NULL
123  *      if everything went OK or the pointer to the last filled triple
124  *      (incomplete one) otherwise. Upon the return chain[i].key contains
125  *      the number of (i+1)-th block in the chain (as it is stored in memory,
126  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
127  *      number (it points into struct inode for i==0 and into the bh->b_data
128  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
129  *      block for i>0 and NULL for i==0. In other words, it holds the block
130  *      numbers of the chain, addresses they were taken from (and where we can
131  *      verify that chain did not change) and buffer_heads hosting these
132  *      numbers.
133  *
134  *      Function stops when it stumbles upon zero pointer (absent block)
135  *              (pointer to last triple returned, *@err == 0)
136  *      or when it gets an IO error reading an indirect block
137  *              (ditto, *@err == -EIO)
138  *      or when it reads all @depth-1 indirect blocks successfully and finds
139  *      the whole chain, all way to the data (returns %NULL, *err == 0).
140  *
141  *      Need to be called with
142  *      down_read(&EXT4_I(inode)->i_data_sem)
143  */
144 static Indirect *ext4_get_branch(struct inode *inode, int depth,
145                                  ext4_lblk_t  *offsets,
146                                  Indirect chain[4], int *err)
147 {
148         struct super_block *sb = inode->i_sb;
149         Indirect *p = chain;
150         struct buffer_head *bh;
151         int ret = -EIO;
152
153         *err = 0;
154         /* i_data is not going away, no lock needed */
155         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
156         if (!p->key)
157                 goto no_block;
158         while (--depth) {
159                 bh = sb_getblk(sb, le32_to_cpu(p->key));
160                 if (unlikely(!bh)) {
161                         ret = -ENOMEM;
162                         goto failure;
163                 }
164
165                 if (!bh_uptodate_or_lock(bh)) {
166                         if (bh_submit_read(bh) < 0) {
167                                 put_bh(bh);
168                                 goto failure;
169                         }
170                         /* validate block references */
171                         if (ext4_check_indirect_blockref(inode, bh)) {
172                                 put_bh(bh);
173                                 goto failure;
174                         }
175                 }
176
177                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
178                 /* Reader: end */
179                 if (!p->key)
180                         goto no_block;
181         }
182         return NULL;
183
184 failure:
185         *err = ret;
186 no_block:
187         return p;
188 }
189
190 /**
191  *      ext4_find_near - find a place for allocation with sufficient locality
192  *      @inode: owner
193  *      @ind: descriptor of indirect block.
194  *
195  *      This function returns the preferred place for block allocation.
196  *      It is used when heuristic for sequential allocation fails.
197  *      Rules are:
198  *        + if there is a block to the left of our position - allocate near it.
199  *        + if pointer will live in indirect block - allocate near that block.
200  *        + if pointer will live in inode - allocate in the same
201  *          cylinder group.
202  *
203  * In the latter case we colour the starting block by the callers PID to
204  * prevent it from clashing with concurrent allocations for a different inode
205  * in the same block group.   The PID is used here so that functionally related
206  * files will be close-by on-disk.
207  *
208  *      Caller must make sure that @ind is valid and will stay that way.
209  */
210 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
211 {
212         struct ext4_inode_info *ei = EXT4_I(inode);
213         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
214         __le32 *p;
215
216         /* Try to find previous block */
217         for (p = ind->p - 1; p >= start; p--) {
218                 if (*p)
219                         return le32_to_cpu(*p);
220         }
221
222         /* No such thing, so let's try location of indirect block */
223         if (ind->bh)
224                 return ind->bh->b_blocknr;
225
226         /*
227          * It is going to be referred to from the inode itself? OK, just put it
228          * into the same cylinder group then.
229          */
230         return ext4_inode_to_goal_block(inode);
231 }
232
233 /**
234  *      ext4_find_goal - find a preferred place for allocation.
235  *      @inode: owner
236  *      @block:  block we want
237  *      @partial: pointer to the last triple within a chain
238  *
239  *      Normally this function find the preferred place for block allocation,
240  *      returns it.
241  *      Because this is only used for non-extent files, we limit the block nr
242  *      to 32 bits.
243  */
244 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
245                                    Indirect *partial)
246 {
247         ext4_fsblk_t goal;
248
249         /*
250          * XXX need to get goal block from mballoc's data structures
251          */
252
253         goal = ext4_find_near(inode, partial);
254         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
255         return goal;
256 }
257
258 /**
259  *      ext4_blks_to_allocate - Look up the block map and count the number
260  *      of direct blocks need to be allocated for the given branch.
261  *
262  *      @branch: chain of indirect blocks
263  *      @k: number of blocks need for indirect blocks
264  *      @blks: number of data blocks to be mapped.
265  *      @blocks_to_boundary:  the offset in the indirect block
266  *
267  *      return the total number of blocks to be allocate, including the
268  *      direct and indirect blocks.
269  */
270 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
271                                  int blocks_to_boundary)
272 {
273         unsigned int count = 0;
274
275         /*
276          * Simple case, [t,d]Indirect block(s) has not allocated yet
277          * then it's clear blocks on that path have not allocated
278          */
279         if (k > 0) {
280                 /* right now we don't handle cross boundary allocation */
281                 if (blks < blocks_to_boundary + 1)
282                         count += blks;
283                 else
284                         count += blocks_to_boundary + 1;
285                 return count;
286         }
287
288         count++;
289         while (count < blks && count <= blocks_to_boundary &&
290                 le32_to_cpu(*(branch[0].p + count)) == 0) {
291                 count++;
292         }
293         return count;
294 }
295
296 /**
297  * ext4_alloc_branch() - allocate and set up a chain of blocks
298  * @handle: handle for this transaction
299  * @ar: structure describing the allocation request
300  * @indirect_blks: number of allocated indirect blocks
301  * @offsets: offsets (in the blocks) to store the pointers to next.
302  * @branch: place to store the chain in.
303  *
304  *      This function allocates blocks, zeroes out all but the last one,
305  *      links them into chain and (if we are synchronous) writes them to disk.
306  *      In other words, it prepares a branch that can be spliced onto the
307  *      inode. It stores the information about that chain in the branch[], in
308  *      the same format as ext4_get_branch() would do. We are calling it after
309  *      we had read the existing part of chain and partial points to the last
310  *      triple of that (one with zero ->key). Upon the exit we have the same
311  *      picture as after the successful ext4_get_block(), except that in one
312  *      place chain is disconnected - *branch->p is still zero (we did not
313  *      set the last link), but branch->key contains the number that should
314  *      be placed into *branch->p to fill that gap.
315  *
316  *      If allocation fails we free all blocks we've allocated (and forget
317  *      their buffer_heads) and return the error value the from failed
318  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
319  *      as described above and return 0.
320  */
321 static int ext4_alloc_branch(handle_t *handle,
322                              struct ext4_allocation_request *ar,
323                              int indirect_blks, ext4_lblk_t *offsets,
324                              Indirect *branch)
325 {
326         struct buffer_head *            bh;
327         ext4_fsblk_t                    b, new_blocks[4];
328         __le32                          *p;
329         int                             i, j, err, len = 1;
330
331         for (i = 0; i <= indirect_blks; i++) {
332                 if (i == indirect_blks) {
333                         new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
334                 } else {
335                         ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
336                                         ar->inode, ar->goal,
337                                         ar->flags & EXT4_MB_DELALLOC_RESERVED,
338                                         NULL, &err);
339                         /* Simplify error cleanup... */
340                         branch[i+1].bh = NULL;
341                 }
342                 if (err) {
343                         i--;
344                         goto failed;
345                 }
346                 branch[i].key = cpu_to_le32(new_blocks[i]);
347                 if (i == 0)
348                         continue;
349
350                 bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
351                 if (unlikely(!bh)) {
352                         err = -ENOMEM;
353                         goto failed;
354                 }
355                 lock_buffer(bh);
356                 BUFFER_TRACE(bh, "call get_create_access");
357                 err = ext4_journal_get_create_access(handle, bh);
358                 if (err) {
359                         unlock_buffer(bh);
360                         goto failed;
361                 }
362
363                 memset(bh->b_data, 0, bh->b_size);
364                 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
365                 b = new_blocks[i];
366
367                 if (i == indirect_blks)
368                         len = ar->len;
369                 for (j = 0; j < len; j++)
370                         *p++ = cpu_to_le32(b++);
371
372                 BUFFER_TRACE(bh, "marking uptodate");
373                 set_buffer_uptodate(bh);
374                 unlock_buffer(bh);
375
376                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
377                 err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
378                 if (err)
379                         goto failed;
380         }
381         return 0;
382 failed:
383         if (i == indirect_blks) {
384                 /* Free data blocks */
385                 ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
386                                  ar->len, 0);
387                 i--;
388         }
389         for (; i >= 0; i--) {
390                 /*
391                  * We want to ext4_forget() only freshly allocated indirect
392                  * blocks. Buffer for new_blocks[i] is at branch[i+1].bh
393                  * (buffer at branch[0].bh is indirect block / inode already
394                  * existing before ext4_alloc_branch() was called). Also
395                  * because blocks are freshly allocated, we don't need to
396                  * revoke them which is why we don't set
397                  * EXT4_FREE_BLOCKS_METADATA.
398                  */
399                 ext4_free_blocks(handle, ar->inode, branch[i+1].bh,
400                                  new_blocks[i], 1,
401                                  branch[i+1].bh ? EXT4_FREE_BLOCKS_FORGET : 0);
402         }
403         return err;
404 }
405
406 /**
407  * ext4_splice_branch() - splice the allocated branch onto inode.
408  * @handle: handle for this transaction
409  * @ar: structure describing the allocation request
410  * @where: location of missing link
411  * @num:   number of indirect blocks we are adding
412  *
413  * This function fills the missing link and does all housekeeping needed in
414  * inode (->i_blocks, etc.). In case of success we end up with the full
415  * chain to new block and return 0.
416  */
417 static int ext4_splice_branch(handle_t *handle,
418                               struct ext4_allocation_request *ar,
419                               Indirect *where, int num)
420 {
421         int i;
422         int err = 0;
423         ext4_fsblk_t current_block;
424
425         /*
426          * If we're splicing into a [td]indirect block (as opposed to the
427          * inode) then we need to get write access to the [td]indirect block
428          * before the splice.
429          */
430         if (where->bh) {
431                 BUFFER_TRACE(where->bh, "get_write_access");
432                 err = ext4_journal_get_write_access(handle, where->bh);
433                 if (err)
434                         goto err_out;
435         }
436         /* That's it */
437
438         *where->p = where->key;
439
440         /*
441          * Update the host buffer_head or inode to point to more just allocated
442          * direct blocks blocks
443          */
444         if (num == 0 && ar->len > 1) {
445                 current_block = le32_to_cpu(where->key) + 1;
446                 for (i = 1; i < ar->len; i++)
447                         *(where->p + i) = cpu_to_le32(current_block++);
448         }
449
450         /* We are done with atomic stuff, now do the rest of housekeeping */
451         /* had we spliced it onto indirect block? */
452         if (where->bh) {
453                 /*
454                  * If we spliced it onto an indirect block, we haven't
455                  * altered the inode.  Note however that if it is being spliced
456                  * onto an indirect block at the very end of the file (the
457                  * file is growing) then we *will* alter the inode to reflect
458                  * the new i_size.  But that is not done here - it is done in
459                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
460                  */
461                 jbd_debug(5, "splicing indirect only\n");
462                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
463                 err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
464                 if (err)
465                         goto err_out;
466         } else {
467                 /*
468                  * OK, we spliced it into the inode itself on a direct block.
469                  */
470                 ext4_mark_inode_dirty(handle, ar->inode);
471                 jbd_debug(5, "splicing direct\n");
472         }
473         return err;
474
475 err_out:
476         for (i = 1; i <= num; i++) {
477                 /*
478                  * branch[i].bh is newly allocated, so there is no
479                  * need to revoke the block, which is why we don't
480                  * need to set EXT4_FREE_BLOCKS_METADATA.
481                  */
482                 ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
483                                  EXT4_FREE_BLOCKS_FORGET);
484         }
485         ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
486                          ar->len, 0);
487
488         return err;
489 }
490
491 /*
492  * The ext4_ind_map_blocks() function handles non-extents inodes
493  * (i.e., using the traditional indirect/double-indirect i_blocks
494  * scheme) for ext4_map_blocks().
495  *
496  * Allocation strategy is simple: if we have to allocate something, we will
497  * have to go the whole way to leaf. So let's do it before attaching anything
498  * to tree, set linkage between the newborn blocks, write them if sync is
499  * required, recheck the path, free and repeat if check fails, otherwise
500  * set the last missing link (that will protect us from any truncate-generated
501  * removals - all blocks on the path are immune now) and possibly force the
502  * write on the parent block.
503  * That has a nice additional property: no special recovery from the failed
504  * allocations is needed - we simply release blocks and do not touch anything
505  * reachable from inode.
506  *
507  * `handle' can be NULL if create == 0.
508  *
509  * return > 0, # of blocks mapped or allocated.
510  * return = 0, if plain lookup failed.
511  * return < 0, error case.
512  *
513  * The ext4_ind_get_blocks() function should be called with
514  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
515  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
516  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
517  * blocks.
518  */
519 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
520                         struct ext4_map_blocks *map,
521                         int flags)
522 {
523         struct ext4_allocation_request ar;
524         int err = -EIO;
525         ext4_lblk_t offsets[4];
526         Indirect chain[4];
527         Indirect *partial;
528         int indirect_blks;
529         int blocks_to_boundary = 0;
530         int depth;
531         int count = 0;
532         ext4_fsblk_t first_block = 0;
533
534         trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
535         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
536         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
537         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
538                                    &blocks_to_boundary);
539
540         if (depth == 0)
541                 goto out;
542
543         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
544
545         /* Simplest case - block found, no allocation needed */
546         if (!partial) {
547                 first_block = le32_to_cpu(chain[depth - 1].key);
548                 count++;
549                 /*map more blocks*/
550                 while (count < map->m_len && count <= blocks_to_boundary) {
551                         ext4_fsblk_t blk;
552
553                         blk = le32_to_cpu(*(chain[depth-1].p + count));
554
555                         if (blk == first_block + count)
556                                 count++;
557                         else
558                                 break;
559                 }
560                 goto got_it;
561         }
562
563         /* Next simple case - plain lookup failed */
564         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
565                 unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
566                 int i;
567
568                 /*
569                  * Count number blocks in a subtree under 'partial'. At each
570                  * level we count number of complete empty subtrees beyond
571                  * current offset and then descend into the subtree only
572                  * partially beyond current offset.
573                  */
574                 count = 0;
575                 for (i = partial - chain + 1; i < depth; i++)
576                         count = count * epb + (epb - offsets[i] - 1);
577                 count++;
578                 /* Fill in size of a hole we found */
579                 map->m_pblk = 0;
580                 map->m_len = min_t(unsigned int, map->m_len, count);
581                 goto cleanup;
582         }
583
584         /* Failed read of indirect block */
585         if (err == -EIO)
586                 goto cleanup;
587
588         /*
589          * Okay, we need to do block allocation.
590         */
591         if (ext4_has_feature_bigalloc(inode->i_sb)) {
592                 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
593                                  "non-extent mapped inodes with bigalloc");
594                 return -EFSCORRUPTED;
595         }
596
597         /* Set up for the direct block allocation */
598         memset(&ar, 0, sizeof(ar));
599         ar.inode = inode;
600         ar.logical = map->m_lblk;
601         if (S_ISREG(inode->i_mode))
602                 ar.flags = EXT4_MB_HINT_DATA;
603         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
604                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
605         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
606                 ar.flags |= EXT4_MB_USE_RESERVED;
607
608         ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
609
610         /* the number of blocks need to allocate for [d,t]indirect blocks */
611         indirect_blks = (chain + depth) - partial - 1;
612
613         /*
614          * Next look up the indirect map to count the totoal number of
615          * direct blocks to allocate for this branch.
616          */
617         ar.len = ext4_blks_to_allocate(partial, indirect_blks,
618                                        map->m_len, blocks_to_boundary);
619
620         /*
621          * Block out ext4_truncate while we alter the tree
622          */
623         err = ext4_alloc_branch(handle, &ar, indirect_blks,
624                                 offsets + (partial - chain), partial);
625
626         /*
627          * The ext4_splice_branch call will free and forget any buffers
628          * on the new chain if there is a failure, but that risks using
629          * up transaction credits, especially for bitmaps where the
630          * credits cannot be returned.  Can we handle this somehow?  We
631          * may need to return -EAGAIN upwards in the worst case.  --sct
632          */
633         if (!err)
634                 err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
635         if (err)
636                 goto cleanup;
637
638         map->m_flags |= EXT4_MAP_NEW;
639
640         ext4_update_inode_fsync_trans(handle, inode, 1);
641         count = ar.len;
642 got_it:
643         map->m_flags |= EXT4_MAP_MAPPED;
644         map->m_pblk = le32_to_cpu(chain[depth-1].key);
645         map->m_len = count;
646         if (count > blocks_to_boundary)
647                 map->m_flags |= EXT4_MAP_BOUNDARY;
648         err = count;
649         /* Clean up and exit */
650         partial = chain + depth - 1;    /* the whole chain */
651 cleanup:
652         while (partial > chain) {
653                 BUFFER_TRACE(partial->bh, "call brelse");
654                 brelse(partial->bh);
655                 partial--;
656         }
657 out:
658         trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
659         return err;
660 }
661
662 /*
663  * Calculate the number of metadata blocks need to reserve
664  * to allocate a new block at @lblocks for non extent file based file
665  */
666 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
667 {
668         struct ext4_inode_info *ei = EXT4_I(inode);
669         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
670         int blk_bits;
671
672         if (lblock < EXT4_NDIR_BLOCKS)
673                 return 0;
674
675         lblock -= EXT4_NDIR_BLOCKS;
676
677         if (ei->i_da_metadata_calc_len &&
678             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
679                 ei->i_da_metadata_calc_len++;
680                 return 0;
681         }
682         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
683         ei->i_da_metadata_calc_len = 1;
684         blk_bits = order_base_2(lblock);
685         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
686 }
687
688 /*
689  * Calculate number of indirect blocks touched by mapping @nrblocks logically
690  * contiguous blocks
691  */
692 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
693 {
694         /*
695          * With N contiguous data blocks, we need at most
696          * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
697          * 2 dindirect blocks, and 1 tindirect block
698          */
699         return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
700 }
701
702 static int ext4_ind_trunc_restart_fn(handle_t *handle, struct inode *inode,
703                                      struct buffer_head *bh, int *dropped)
704 {
705         int err;
706
707         if (bh) {
708                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
709                 err = ext4_handle_dirty_metadata(handle, inode, bh);
710                 if (unlikely(err))
711                         return err;
712         }
713         err = ext4_mark_inode_dirty(handle, inode);
714         if (unlikely(err))
715                 return err;
716         /*
717          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
718          * moment, get_block can be called only for blocks inside i_size since
719          * page cache has been already dropped and writes are blocked by
720          * i_mutex. So we can safely drop the i_data_sem here.
721          */
722         BUG_ON(EXT4_JOURNAL(inode) == NULL);
723         ext4_discard_preallocations(inode);
724         up_write(&EXT4_I(inode)->i_data_sem);
725         *dropped = 1;
726         return 0;
727 }
728
729 /*
730  * Truncate transactions can be complex and absolutely huge.  So we need to
731  * be able to restart the transaction at a conventient checkpoint to make
732  * sure we don't overflow the journal.
733  *
734  * Try to extend this transaction for the purposes of truncation.  If
735  * extend fails, we restart transaction.
736  */
737 static int ext4_ind_truncate_ensure_credits(handle_t *handle,
738                                             struct inode *inode,
739                                             struct buffer_head *bh,
740                                             int revoke_creds)
741 {
742         int ret;
743         int dropped = 0;
744
745         ret = ext4_journal_ensure_credits_fn(handle, EXT4_RESERVE_TRANS_BLOCKS,
746                         ext4_blocks_for_truncate(inode), revoke_creds,
747                         ext4_ind_trunc_restart_fn(handle, inode, bh, &dropped));
748         if (dropped)
749                 down_write(&EXT4_I(inode)->i_data_sem);
750         if (ret <= 0)
751                 return ret;
752         if (bh) {
753                 BUFFER_TRACE(bh, "retaking write access");
754                 ret = ext4_journal_get_write_access(handle, bh);
755                 if (unlikely(ret))
756                         return ret;
757         }
758         return 0;
759 }
760
761 /*
762  * Probably it should be a library function... search for first non-zero word
763  * or memcmp with zero_page, whatever is better for particular architecture.
764  * Linus?
765  */
766 static inline int all_zeroes(__le32 *p, __le32 *q)
767 {
768         while (p < q)
769                 if (*p++)
770                         return 0;
771         return 1;
772 }
773
774 /**
775  *      ext4_find_shared - find the indirect blocks for partial truncation.
776  *      @inode:   inode in question
777  *      @depth:   depth of the affected branch
778  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
779  *      @chain:   place to store the pointers to partial indirect blocks
780  *      @top:     place to the (detached) top of branch
781  *
782  *      This is a helper function used by ext4_truncate().
783  *
784  *      When we do truncate() we may have to clean the ends of several
785  *      indirect blocks but leave the blocks themselves alive. Block is
786  *      partially truncated if some data below the new i_size is referred
787  *      from it (and it is on the path to the first completely truncated
788  *      data block, indeed).  We have to free the top of that path along
789  *      with everything to the right of the path. Since no allocation
790  *      past the truncation point is possible until ext4_truncate()
791  *      finishes, we may safely do the latter, but top of branch may
792  *      require special attention - pageout below the truncation point
793  *      might try to populate it.
794  *
795  *      We atomically detach the top of branch from the tree, store the
796  *      block number of its root in *@top, pointers to buffer_heads of
797  *      partially truncated blocks - in @chain[].bh and pointers to
798  *      their last elements that should not be removed - in
799  *      @chain[].p. Return value is the pointer to last filled element
800  *      of @chain.
801  *
802  *      The work left to caller to do the actual freeing of subtrees:
803  *              a) free the subtree starting from *@top
804  *              b) free the subtrees whose roots are stored in
805  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
806  *              c) free the subtrees growing from the inode past the @chain[0].
807  *                      (no partially truncated stuff there).  */
808
809 static Indirect *ext4_find_shared(struct inode *inode, int depth,
810                                   ext4_lblk_t offsets[4], Indirect chain[4],
811                                   __le32 *top)
812 {
813         Indirect *partial, *p;
814         int k, err;
815
816         *top = 0;
817         /* Make k index the deepest non-null offset + 1 */
818         for (k = depth; k > 1 && !offsets[k-1]; k--)
819                 ;
820         partial = ext4_get_branch(inode, k, offsets, chain, &err);
821         /* Writer: pointers */
822         if (!partial)
823                 partial = chain + k-1;
824         /*
825          * If the branch acquired continuation since we've looked at it -
826          * fine, it should all survive and (new) top doesn't belong to us.
827          */
828         if (!partial->key && *partial->p)
829                 /* Writer: end */
830                 goto no_top;
831         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
832                 ;
833         /*
834          * OK, we've found the last block that must survive. The rest of our
835          * branch should be detached before unlocking. However, if that rest
836          * of branch is all ours and does not grow immediately from the inode
837          * it's easier to cheat and just decrement partial->p.
838          */
839         if (p == chain + k - 1 && p > chain) {
840                 p->p--;
841         } else {
842                 *top = *p->p;
843                 /* Nope, don't do this in ext4.  Must leave the tree intact */
844 #if 0
845                 *p->p = 0;
846 #endif
847         }
848         /* Writer: end */
849
850         while (partial > p) {
851                 brelse(partial->bh);
852                 partial--;
853         }
854 no_top:
855         return partial;
856 }
857
858 /*
859  * Zero a number of block pointers in either an inode or an indirect block.
860  * If we restart the transaction we must again get write access to the
861  * indirect block for further modification.
862  *
863  * We release `count' blocks on disk, but (last - first) may be greater
864  * than `count' because there can be holes in there.
865  *
866  * Return 0 on success, 1 on invalid block range
867  * and < 0 on fatal error.
868  */
869 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
870                              struct buffer_head *bh,
871                              ext4_fsblk_t block_to_free,
872                              unsigned long count, __le32 *first,
873                              __le32 *last)
874 {
875         __le32 *p;
876         int     flags = EXT4_FREE_BLOCKS_VALIDATED;
877         int     err;
878
879         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
880             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
881                 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
882         else if (ext4_should_journal_data(inode))
883                 flags |= EXT4_FREE_BLOCKS_FORGET;
884
885         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
886                                    count)) {
887                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
888                                  "blocks %llu len %lu",
889                                  (unsigned long long) block_to_free, count);
890                 return 1;
891         }
892
893         err = ext4_ind_truncate_ensure_credits(handle, inode, bh,
894                                 ext4_free_data_revoke_credits(inode, count));
895         if (err < 0)
896                 goto out_err;
897
898         for (p = first; p < last; p++)
899                 *p = 0;
900
901         ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
902         return 0;
903 out_err:
904         ext4_std_error(inode->i_sb, err);
905         return err;
906 }
907
908 /**
909  * ext4_free_data - free a list of data blocks
910  * @handle:     handle for this transaction
911  * @inode:      inode we are dealing with
912  * @this_bh:    indirect buffer_head which contains *@first and *@last
913  * @first:      array of block numbers
914  * @last:       points immediately past the end of array
915  *
916  * We are freeing all blocks referred from that array (numbers are stored as
917  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
918  *
919  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
920  * blocks are contiguous then releasing them at one time will only affect one
921  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
922  * actually use a lot of journal space.
923  *
924  * @this_bh will be %NULL if @first and @last point into the inode's direct
925  * block pointers.
926  */
927 static void ext4_free_data(handle_t *handle, struct inode *inode,
928                            struct buffer_head *this_bh,
929                            __le32 *first, __le32 *last)
930 {
931         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
932         unsigned long count = 0;            /* Number of blocks in the run */
933         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
934                                                corresponding to
935                                                block_to_free */
936         ext4_fsblk_t nr;                    /* Current block # */
937         __le32 *p;                          /* Pointer into inode/ind
938                                                for current block */
939         int err = 0;
940
941         if (this_bh) {                          /* For indirect block */
942                 BUFFER_TRACE(this_bh, "get_write_access");
943                 err = ext4_journal_get_write_access(handle, this_bh);
944                 /* Important: if we can't update the indirect pointers
945                  * to the blocks, we can't free them. */
946                 if (err)
947                         return;
948         }
949
950         for (p = first; p < last; p++) {
951                 nr = le32_to_cpu(*p);
952                 if (nr) {
953                         /* accumulate blocks to free if they're contiguous */
954                         if (count == 0) {
955                                 block_to_free = nr;
956                                 block_to_free_p = p;
957                                 count = 1;
958                         } else if (nr == block_to_free + count) {
959                                 count++;
960                         } else {
961                                 err = ext4_clear_blocks(handle, inode, this_bh,
962                                                         block_to_free, count,
963                                                         block_to_free_p, p);
964                                 if (err)
965                                         break;
966                                 block_to_free = nr;
967                                 block_to_free_p = p;
968                                 count = 1;
969                         }
970                 }
971         }
972
973         if (!err && count > 0)
974                 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
975                                         count, block_to_free_p, p);
976         if (err < 0)
977                 /* fatal error */
978                 return;
979
980         if (this_bh) {
981                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
982
983                 /*
984                  * The buffer head should have an attached journal head at this
985                  * point. However, if the data is corrupted and an indirect
986                  * block pointed to itself, it would have been detached when
987                  * the block was cleared. Check for this instead of OOPSing.
988                  */
989                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
990                         ext4_handle_dirty_metadata(handle, inode, this_bh);
991                 else
992                         EXT4_ERROR_INODE(inode,
993                                          "circular indirect block detected at "
994                                          "block %llu",
995                                 (unsigned long long) this_bh->b_blocknr);
996         }
997 }
998
999 /**
1000  *      ext4_free_branches - free an array of branches
1001  *      @handle: JBD handle for this transaction
1002  *      @inode: inode we are dealing with
1003  *      @parent_bh: the buffer_head which contains *@first and *@last
1004  *      @first: array of block numbers
1005  *      @last:  pointer immediately past the end of array
1006  *      @depth: depth of the branches to free
1007  *
1008  *      We are freeing all blocks referred from these branches (numbers are
1009  *      stored as little-endian 32-bit) and updating @inode->i_blocks
1010  *      appropriately.
1011  */
1012 static void ext4_free_branches(handle_t *handle, struct inode *inode,
1013                                struct buffer_head *parent_bh,
1014                                __le32 *first, __le32 *last, int depth)
1015 {
1016         ext4_fsblk_t nr;
1017         __le32 *p;
1018
1019         if (ext4_handle_is_aborted(handle))
1020                 return;
1021
1022         if (depth--) {
1023                 struct buffer_head *bh;
1024                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1025                 p = last;
1026                 while (--p >= first) {
1027                         nr = le32_to_cpu(*p);
1028                         if (!nr)
1029                                 continue;               /* A hole */
1030
1031                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1032                                                    nr, 1)) {
1033                                 EXT4_ERROR_INODE(inode,
1034                                                  "invalid indirect mapped "
1035                                                  "block %lu (level %d)",
1036                                                  (unsigned long) nr, depth);
1037                                 break;
1038                         }
1039
1040                         /* Go read the buffer for the next level down */
1041                         bh = sb_bread(inode->i_sb, nr);
1042
1043                         /*
1044                          * A read failure? Report error and clear slot
1045                          * (should be rare).
1046                          */
1047                         if (!bh) {
1048                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
1049                                                        "Read failure");
1050                                 continue;
1051                         }
1052
1053                         /* This zaps the entire block.  Bottom up. */
1054                         BUFFER_TRACE(bh, "free child branches");
1055                         ext4_free_branches(handle, inode, bh,
1056                                         (__le32 *) bh->b_data,
1057                                         (__le32 *) bh->b_data + addr_per_block,
1058                                         depth);
1059                         brelse(bh);
1060
1061                         /*
1062                          * Everything below this this pointer has been
1063                          * released.  Now let this top-of-subtree go.
1064                          *
1065                          * We want the freeing of this indirect block to be
1066                          * atomic in the journal with the updating of the
1067                          * bitmap block which owns it.  So make some room in
1068                          * the journal.
1069                          *
1070                          * We zero the parent pointer *after* freeing its
1071                          * pointee in the bitmaps, so if extend_transaction()
1072                          * for some reason fails to put the bitmap changes and
1073                          * the release into the same transaction, recovery
1074                          * will merely complain about releasing a free block,
1075                          * rather than leaking blocks.
1076                          */
1077                         if (ext4_handle_is_aborted(handle))
1078                                 return;
1079                         if (ext4_ind_truncate_ensure_credits(handle, inode,
1080                                         NULL,
1081                                         ext4_free_metadata_revoke_credits(
1082                                                         inode->i_sb, 1)) < 0)
1083                                 return;
1084
1085                         /*
1086                          * The forget flag here is critical because if
1087                          * we are journaling (and not doing data
1088                          * journaling), we have to make sure a revoke
1089                          * record is written to prevent the journal
1090                          * replay from overwriting the (former)
1091                          * indirect block if it gets reallocated as a
1092                          * data block.  This must happen in the same
1093                          * transaction where the data blocks are
1094                          * actually freed.
1095                          */
1096                         ext4_free_blocks(handle, inode, NULL, nr, 1,
1097                                          EXT4_FREE_BLOCKS_METADATA|
1098                                          EXT4_FREE_BLOCKS_FORGET);
1099
1100                         if (parent_bh) {
1101                                 /*
1102                                  * The block which we have just freed is
1103                                  * pointed to by an indirect block: journal it
1104                                  */
1105                                 BUFFER_TRACE(parent_bh, "get_write_access");
1106                                 if (!ext4_journal_get_write_access(handle,
1107                                                                    parent_bh)){
1108                                         *p = 0;
1109                                         BUFFER_TRACE(parent_bh,
1110                                         "call ext4_handle_dirty_metadata");
1111                                         ext4_handle_dirty_metadata(handle,
1112                                                                    inode,
1113                                                                    parent_bh);
1114                                 }
1115                         }
1116                 }
1117         } else {
1118                 /* We have reached the bottom of the tree. */
1119                 BUFFER_TRACE(parent_bh, "free data blocks");
1120                 ext4_free_data(handle, inode, parent_bh, first, last);
1121         }
1122 }
1123
1124 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1125 {
1126         struct ext4_inode_info *ei = EXT4_I(inode);
1127         __le32 *i_data = ei->i_data;
1128         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1129         ext4_lblk_t offsets[4];
1130         Indirect chain[4];
1131         Indirect *partial;
1132         __le32 nr = 0;
1133         int n = 0;
1134         ext4_lblk_t last_block, max_block;
1135         unsigned blocksize = inode->i_sb->s_blocksize;
1136
1137         last_block = (inode->i_size + blocksize-1)
1138                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1139         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1140                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1141
1142         if (last_block != max_block) {
1143                 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1144                 if (n == 0)
1145                         return;
1146         }
1147
1148         ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1149
1150         /*
1151          * The orphan list entry will now protect us from any crash which
1152          * occurs before the truncate completes, so it is now safe to propagate
1153          * the new, shorter inode size (held for now in i_size) into the
1154          * on-disk inode. We do this via i_disksize, which is the value which
1155          * ext4 *really* writes onto the disk inode.
1156          */
1157         ei->i_disksize = inode->i_size;
1158
1159         if (last_block == max_block) {
1160                 /*
1161                  * It is unnecessary to free any data blocks if last_block is
1162                  * equal to the indirect block limit.
1163                  */
1164                 return;
1165         } else if (n == 1) {            /* direct blocks */
1166                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1167                                i_data + EXT4_NDIR_BLOCKS);
1168                 goto do_indirects;
1169         }
1170
1171         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1172         /* Kill the top of shared branch (not detached) */
1173         if (nr) {
1174                 if (partial == chain) {
1175                         /* Shared branch grows from the inode */
1176                         ext4_free_branches(handle, inode, NULL,
1177                                            &nr, &nr+1, (chain+n-1) - partial);
1178                         *partial->p = 0;
1179                         /*
1180                          * We mark the inode dirty prior to restart,
1181                          * and prior to stop.  No need for it here.
1182                          */
1183                 } else {
1184                         /* Shared branch grows from an indirect block */
1185                         BUFFER_TRACE(partial->bh, "get_write_access");
1186                         ext4_free_branches(handle, inode, partial->bh,
1187                                         partial->p,
1188                                         partial->p+1, (chain+n-1) - partial);
1189                 }
1190         }
1191         /* Clear the ends of indirect blocks on the shared branch */
1192         while (partial > chain) {
1193                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1194                                    (__le32*)partial->bh->b_data+addr_per_block,
1195                                    (chain+n-1) - partial);
1196                 BUFFER_TRACE(partial->bh, "call brelse");
1197                 brelse(partial->bh);
1198                 partial--;
1199         }
1200 do_indirects:
1201         /* Kill the remaining (whole) subtrees */
1202         switch (offsets[0]) {
1203         default:
1204                 nr = i_data[EXT4_IND_BLOCK];
1205                 if (nr) {
1206                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1207                         i_data[EXT4_IND_BLOCK] = 0;
1208                 }
1209                 /* fall through */
1210         case EXT4_IND_BLOCK:
1211                 nr = i_data[EXT4_DIND_BLOCK];
1212                 if (nr) {
1213                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1214                         i_data[EXT4_DIND_BLOCK] = 0;
1215                 }
1216                 /* fall through */
1217         case EXT4_DIND_BLOCK:
1218                 nr = i_data[EXT4_TIND_BLOCK];
1219                 if (nr) {
1220                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1221                         i_data[EXT4_TIND_BLOCK] = 0;
1222                 }
1223                 /* fall through */
1224         case EXT4_TIND_BLOCK:
1225                 ;
1226         }
1227 }
1228
1229 /**
1230  *      ext4_ind_remove_space - remove space from the range
1231  *      @handle: JBD handle for this transaction
1232  *      @inode: inode we are dealing with
1233  *      @start: First block to remove
1234  *      @end:   One block after the last block to remove (exclusive)
1235  *
1236  *      Free the blocks in the defined range (end is exclusive endpoint of
1237  *      range). This is used by ext4_punch_hole().
1238  */
1239 int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1240                           ext4_lblk_t start, ext4_lblk_t end)
1241 {
1242         struct ext4_inode_info *ei = EXT4_I(inode);
1243         __le32 *i_data = ei->i_data;
1244         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1245         ext4_lblk_t offsets[4], offsets2[4];
1246         Indirect chain[4], chain2[4];
1247         Indirect *partial, *partial2;
1248         Indirect *p = NULL, *p2 = NULL;
1249         ext4_lblk_t max_block;
1250         __le32 nr = 0, nr2 = 0;
1251         int n = 0, n2 = 0;
1252         unsigned blocksize = inode->i_sb->s_blocksize;
1253
1254         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1255                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1256         if (end >= max_block)
1257                 end = max_block;
1258         if ((start >= end) || (start > max_block))
1259                 return 0;
1260
1261         n = ext4_block_to_path(inode, start, offsets, NULL);
1262         n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1263
1264         BUG_ON(n > n2);
1265
1266         if ((n == 1) && (n == n2)) {
1267                 /* We're punching only within direct block range */
1268                 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1269                                i_data + offsets2[0]);
1270                 return 0;
1271         } else if (n2 > n) {
1272                 /*
1273                  * Start and end are on a different levels so we're going to
1274                  * free partial block at start, and partial block at end of
1275                  * the range. If there are some levels in between then
1276                  * do_indirects label will take care of that.
1277                  */
1278
1279                 if (n == 1) {
1280                         /*
1281                          * Start is at the direct block level, free
1282                          * everything to the end of the level.
1283                          */
1284                         ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1285                                        i_data + EXT4_NDIR_BLOCKS);
1286                         goto end_range;
1287                 }
1288
1289
1290                 partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1291                 if (nr) {
1292                         if (partial == chain) {
1293                                 /* Shared branch grows from the inode */
1294                                 ext4_free_branches(handle, inode, NULL,
1295                                            &nr, &nr+1, (chain+n-1) - partial);
1296                                 *partial->p = 0;
1297                         } else {
1298                                 /* Shared branch grows from an indirect block */
1299                                 BUFFER_TRACE(partial->bh, "get_write_access");
1300                                 ext4_free_branches(handle, inode, partial->bh,
1301                                         partial->p,
1302                                         partial->p+1, (chain+n-1) - partial);
1303                         }
1304                 }
1305
1306                 /*
1307                  * Clear the ends of indirect blocks on the shared branch
1308                  * at the start of the range
1309                  */
1310                 while (partial > chain) {
1311                         ext4_free_branches(handle, inode, partial->bh,
1312                                 partial->p + 1,
1313                                 (__le32 *)partial->bh->b_data+addr_per_block,
1314                                 (chain+n-1) - partial);
1315                         partial--;
1316                 }
1317
1318 end_range:
1319                 partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1320                 if (nr2) {
1321                         if (partial2 == chain2) {
1322                                 /*
1323                                  * Remember, end is exclusive so here we're at
1324                                  * the start of the next level we're not going
1325                                  * to free. Everything was covered by the start
1326                                  * of the range.
1327                                  */
1328                                 goto do_indirects;
1329                         }
1330                 } else {
1331                         /*
1332                          * ext4_find_shared returns Indirect structure which
1333                          * points to the last element which should not be
1334                          * removed by truncate. But this is end of the range
1335                          * in punch_hole so we need to point to the next element
1336                          */
1337                         partial2->p++;
1338                 }
1339
1340                 /*
1341                  * Clear the ends of indirect blocks on the shared branch
1342                  * at the end of the range
1343                  */
1344                 while (partial2 > chain2) {
1345                         ext4_free_branches(handle, inode, partial2->bh,
1346                                            (__le32 *)partial2->bh->b_data,
1347                                            partial2->p,
1348                                            (chain2+n2-1) - partial2);
1349                         partial2--;
1350                 }
1351                 goto do_indirects;
1352         }
1353
1354         /* Punch happened within the same level (n == n2) */
1355         partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1356         partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1357
1358         /* Free top, but only if partial2 isn't its subtree. */
1359         if (nr) {
1360                 int level = min(partial - chain, partial2 - chain2);
1361                 int i;
1362                 int subtree = 1;
1363
1364                 for (i = 0; i <= level; i++) {
1365                         if (offsets[i] != offsets2[i]) {
1366                                 subtree = 0;
1367                                 break;
1368                         }
1369                 }
1370
1371                 if (!subtree) {
1372                         if (partial == chain) {
1373                                 /* Shared branch grows from the inode */
1374                                 ext4_free_branches(handle, inode, NULL,
1375                                                    &nr, &nr+1,
1376                                                    (chain+n-1) - partial);
1377                                 *partial->p = 0;
1378                         } else {
1379                                 /* Shared branch grows from an indirect block */
1380                                 BUFFER_TRACE(partial->bh, "get_write_access");
1381                                 ext4_free_branches(handle, inode, partial->bh,
1382                                                    partial->p,
1383                                                    partial->p+1,
1384                                                    (chain+n-1) - partial);
1385                         }
1386                 }
1387         }
1388
1389         if (!nr2) {
1390                 /*
1391                  * ext4_find_shared returns Indirect structure which
1392                  * points to the last element which should not be
1393                  * removed by truncate. But this is end of the range
1394                  * in punch_hole so we need to point to the next element
1395                  */
1396                 partial2->p++;
1397         }
1398
1399         while (partial > chain || partial2 > chain2) {
1400                 int depth = (chain+n-1) - partial;
1401                 int depth2 = (chain2+n2-1) - partial2;
1402
1403                 if (partial > chain && partial2 > chain2 &&
1404                     partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1405                         /*
1406                          * We've converged on the same block. Clear the range,
1407                          * then we're done.
1408                          */
1409                         ext4_free_branches(handle, inode, partial->bh,
1410                                            partial->p + 1,
1411                                            partial2->p,
1412                                            (chain+n-1) - partial);
1413                         goto cleanup;
1414                 }
1415
1416                 /*
1417                  * The start and end partial branches may not be at the same
1418                  * level even though the punch happened within one level. So, we
1419                  * give them a chance to arrive at the same level, then walk
1420                  * them in step with each other until we converge on the same
1421                  * block.
1422                  */
1423                 if (partial > chain && depth <= depth2) {
1424                         ext4_free_branches(handle, inode, partial->bh,
1425                                            partial->p + 1,
1426                                            (__le32 *)partial->bh->b_data+addr_per_block,
1427                                            (chain+n-1) - partial);
1428                         partial--;
1429                 }
1430                 if (partial2 > chain2 && depth2 <= depth) {
1431                         ext4_free_branches(handle, inode, partial2->bh,
1432                                            (__le32 *)partial2->bh->b_data,
1433                                            partial2->p,
1434                                            (chain2+n2-1) - partial2);
1435                         partial2--;
1436                 }
1437         }
1438
1439 cleanup:
1440         while (p && p > chain) {
1441                 BUFFER_TRACE(p->bh, "call brelse");
1442                 brelse(p->bh);
1443                 p--;
1444         }
1445         while (p2 && p2 > chain2) {
1446                 BUFFER_TRACE(p2->bh, "call brelse");
1447                 brelse(p2->bh);
1448                 p2--;
1449         }
1450         return 0;
1451
1452 do_indirects:
1453         /* Kill the remaining (whole) subtrees */
1454         switch (offsets[0]) {
1455         default:
1456                 if (++n >= n2)
1457                         break;
1458                 nr = i_data[EXT4_IND_BLOCK];
1459                 if (nr) {
1460                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1461                         i_data[EXT4_IND_BLOCK] = 0;
1462                 }
1463                 /* fall through */
1464         case EXT4_IND_BLOCK:
1465                 if (++n >= n2)
1466                         break;
1467                 nr = i_data[EXT4_DIND_BLOCK];
1468                 if (nr) {
1469                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1470                         i_data[EXT4_DIND_BLOCK] = 0;
1471                 }
1472                 /* fall through */
1473         case EXT4_DIND_BLOCK:
1474                 if (++n >= n2)
1475                         break;
1476                 nr = i_data[EXT4_TIND_BLOCK];
1477                 if (nr) {
1478                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1479                         i_data[EXT4_TIND_BLOCK] = 0;
1480                 }
1481                 /* fall through */
1482         case EXT4_TIND_BLOCK:
1483                 ;
1484         }
1485         goto cleanup;
1486 }