GNU Linux-libre 4.9.308-gnu1
[releases.git] / fs / ext4 / indirect.c
1 /*
2  *  linux/fs/ext4/indirect.c
3  *
4  *  from
5  *
6  *  linux/fs/ext4/inode.c
7  *
8  * Copyright (C) 1992, 1993, 1994, 1995
9  * Remy Card (card@masi.ibp.fr)
10  * Laboratoire MASI - Institut Blaise Pascal
11  * Universite Pierre et Marie Curie (Paris VI)
12  *
13  *  from
14  *
15  *  linux/fs/minix/inode.c
16  *
17  *  Copyright (C) 1991, 1992  Linus Torvalds
18  *
19  *  Goal-directed block allocation by Stephen Tweedie
20  *      (sct@redhat.com), 1993, 1998
21  */
22
23 #include "ext4_jbd2.h"
24 #include "truncate.h"
25 #include <linux/dax.h>
26 #include <linux/uio.h>
27
28 #include <trace/events/ext4.h>
29
30 typedef struct {
31         __le32  *p;
32         __le32  key;
33         struct buffer_head *bh;
34 } Indirect;
35
36 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
37 {
38         p->key = *(p->p = v);
39         p->bh = bh;
40 }
41
42 /**
43  *      ext4_block_to_path - parse the block number into array of offsets
44  *      @inode: inode in question (we are only interested in its superblock)
45  *      @i_block: block number to be parsed
46  *      @offsets: array to store the offsets in
47  *      @boundary: set this non-zero if the referred-to block is likely to be
48  *             followed (on disk) by an indirect block.
49  *
50  *      To store the locations of file's data ext4 uses a data structure common
51  *      for UNIX filesystems - tree of pointers anchored in the inode, with
52  *      data blocks at leaves and indirect blocks in intermediate nodes.
53  *      This function translates the block number into path in that tree -
54  *      return value is the path length and @offsets[n] is the offset of
55  *      pointer to (n+1)th node in the nth one. If @block is out of range
56  *      (negative or too large) warning is printed and zero returned.
57  *
58  *      Note: function doesn't find node addresses, so no IO is needed. All
59  *      we need to know is the capacity of indirect blocks (taken from the
60  *      inode->i_sb).
61  */
62
63 /*
64  * Portability note: the last comparison (check that we fit into triple
65  * indirect block) is spelled differently, because otherwise on an
66  * architecture with 32-bit longs and 8Kb pages we might get into trouble
67  * if our filesystem had 8Kb blocks. We might use long long, but that would
68  * kill us on x86. Oh, well, at least the sign propagation does not matter -
69  * i_block would have to be negative in the very beginning, so we would not
70  * get there at all.
71  */
72
73 static int ext4_block_to_path(struct inode *inode,
74                               ext4_lblk_t i_block,
75                               ext4_lblk_t offsets[4], int *boundary)
76 {
77         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
78         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
79         const long direct_blocks = EXT4_NDIR_BLOCKS,
80                 indirect_blocks = ptrs,
81                 double_blocks = (1 << (ptrs_bits * 2));
82         int n = 0;
83         int final = 0;
84
85         if (i_block < direct_blocks) {
86                 offsets[n++] = i_block;
87                 final = direct_blocks;
88         } else if ((i_block -= direct_blocks) < indirect_blocks) {
89                 offsets[n++] = EXT4_IND_BLOCK;
90                 offsets[n++] = i_block;
91                 final = ptrs;
92         } else if ((i_block -= indirect_blocks) < double_blocks) {
93                 offsets[n++] = EXT4_DIND_BLOCK;
94                 offsets[n++] = i_block >> ptrs_bits;
95                 offsets[n++] = i_block & (ptrs - 1);
96                 final = ptrs;
97         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
98                 offsets[n++] = EXT4_TIND_BLOCK;
99                 offsets[n++] = i_block >> (ptrs_bits * 2);
100                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
101                 offsets[n++] = i_block & (ptrs - 1);
102                 final = ptrs;
103         } else {
104                 ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
105                              i_block + direct_blocks +
106                              indirect_blocks + double_blocks, inode->i_ino);
107         }
108         if (boundary)
109                 *boundary = final - 1 - (i_block & (ptrs - 1));
110         return n;
111 }
112
113 /**
114  *      ext4_get_branch - read the chain of indirect blocks leading to data
115  *      @inode: inode in question
116  *      @depth: depth of the chain (1 - direct pointer, etc.)
117  *      @offsets: offsets of pointers in inode/indirect blocks
118  *      @chain: place to store the result
119  *      @err: here we store the error value
120  *
121  *      Function fills the array of triples <key, p, bh> and returns %NULL
122  *      if everything went OK or the pointer to the last filled triple
123  *      (incomplete one) otherwise. Upon the return chain[i].key contains
124  *      the number of (i+1)-th block in the chain (as it is stored in memory,
125  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
126  *      number (it points into struct inode for i==0 and into the bh->b_data
127  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
128  *      block for i>0 and NULL for i==0. In other words, it holds the block
129  *      numbers of the chain, addresses they were taken from (and where we can
130  *      verify that chain did not change) and buffer_heads hosting these
131  *      numbers.
132  *
133  *      Function stops when it stumbles upon zero pointer (absent block)
134  *              (pointer to last triple returned, *@err == 0)
135  *      or when it gets an IO error reading an indirect block
136  *              (ditto, *@err == -EIO)
137  *      or when it reads all @depth-1 indirect blocks successfully and finds
138  *      the whole chain, all way to the data (returns %NULL, *err == 0).
139  *
140  *      Need to be called with
141  *      down_read(&EXT4_I(inode)->i_data_sem)
142  */
143 static Indirect *ext4_get_branch(struct inode *inode, int depth,
144                                  ext4_lblk_t  *offsets,
145                                  Indirect chain[4], int *err)
146 {
147         struct super_block *sb = inode->i_sb;
148         Indirect *p = chain;
149         struct buffer_head *bh;
150         int ret = -EIO;
151
152         *err = 0;
153         /* i_data is not going away, no lock needed */
154         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
155         if (!p->key)
156                 goto no_block;
157         while (--depth) {
158                 bh = sb_getblk(sb, le32_to_cpu(p->key));
159                 if (unlikely(!bh)) {
160                         ret = -ENOMEM;
161                         goto failure;
162                 }
163
164                 if (!bh_uptodate_or_lock(bh)) {
165                         if (bh_submit_read(bh) < 0) {
166                                 put_bh(bh);
167                                 goto failure;
168                         }
169                         /* validate block references */
170                         if (ext4_check_indirect_blockref(inode, bh)) {
171                                 put_bh(bh);
172                                 goto failure;
173                         }
174                 }
175
176                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
177                 /* Reader: end */
178                 if (!p->key)
179                         goto no_block;
180         }
181         return NULL;
182
183 failure:
184         *err = ret;
185 no_block:
186         return p;
187 }
188
189 /**
190  *      ext4_find_near - find a place for allocation with sufficient locality
191  *      @inode: owner
192  *      @ind: descriptor of indirect block.
193  *
194  *      This function returns the preferred place for block allocation.
195  *      It is used when heuristic for sequential allocation fails.
196  *      Rules are:
197  *        + if there is a block to the left of our position - allocate near it.
198  *        + if pointer will live in indirect block - allocate near that block.
199  *        + if pointer will live in inode - allocate in the same
200  *          cylinder group.
201  *
202  * In the latter case we colour the starting block by the callers PID to
203  * prevent it from clashing with concurrent allocations for a different inode
204  * in the same block group.   The PID is used here so that functionally related
205  * files will be close-by on-disk.
206  *
207  *      Caller must make sure that @ind is valid and will stay that way.
208  */
209 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
210 {
211         struct ext4_inode_info *ei = EXT4_I(inode);
212         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
213         __le32 *p;
214
215         /* Try to find previous block */
216         for (p = ind->p - 1; p >= start; p--) {
217                 if (*p)
218                         return le32_to_cpu(*p);
219         }
220
221         /* No such thing, so let's try location of indirect block */
222         if (ind->bh)
223                 return ind->bh->b_blocknr;
224
225         /*
226          * It is going to be referred to from the inode itself? OK, just put it
227          * into the same cylinder group then.
228          */
229         return ext4_inode_to_goal_block(inode);
230 }
231
232 /**
233  *      ext4_find_goal - find a preferred place for allocation.
234  *      @inode: owner
235  *      @block:  block we want
236  *      @partial: pointer to the last triple within a chain
237  *
238  *      Normally this function find the preferred place for block allocation,
239  *      returns it.
240  *      Because this is only used for non-extent files, we limit the block nr
241  *      to 32 bits.
242  */
243 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
244                                    Indirect *partial)
245 {
246         ext4_fsblk_t goal;
247
248         /*
249          * XXX need to get goal block from mballoc's data structures
250          */
251
252         goal = ext4_find_near(inode, partial);
253         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
254         return goal;
255 }
256
257 /**
258  *      ext4_blks_to_allocate - Look up the block map and count the number
259  *      of direct blocks need to be allocated for the given branch.
260  *
261  *      @branch: chain of indirect blocks
262  *      @k: number of blocks need for indirect blocks
263  *      @blks: number of data blocks to be mapped.
264  *      @blocks_to_boundary:  the offset in the indirect block
265  *
266  *      return the total number of blocks to be allocate, including the
267  *      direct and indirect blocks.
268  */
269 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
270                                  int blocks_to_boundary)
271 {
272         unsigned int count = 0;
273
274         /*
275          * Simple case, [t,d]Indirect block(s) has not allocated yet
276          * then it's clear blocks on that path have not allocated
277          */
278         if (k > 0) {
279                 /* right now we don't handle cross boundary allocation */
280                 if (blks < blocks_to_boundary + 1)
281                         count += blks;
282                 else
283                         count += blocks_to_boundary + 1;
284                 return count;
285         }
286
287         count++;
288         while (count < blks && count <= blocks_to_boundary &&
289                 le32_to_cpu(*(branch[0].p + count)) == 0) {
290                 count++;
291         }
292         return count;
293 }
294
295 /**
296  *      ext4_alloc_branch - allocate and set up a chain of blocks.
297  *      @handle: handle for this transaction
298  *      @inode: owner
299  *      @indirect_blks: number of allocated indirect blocks
300  *      @blks: number of allocated direct blocks
301  *      @goal: preferred place for allocation
302  *      @offsets: offsets (in the blocks) to store the pointers to next.
303  *      @branch: place to store the chain in.
304  *
305  *      This function allocates blocks, zeroes out all but the last one,
306  *      links them into chain and (if we are synchronous) writes them to disk.
307  *      In other words, it prepares a branch that can be spliced onto the
308  *      inode. It stores the information about that chain in the branch[], in
309  *      the same format as ext4_get_branch() would do. We are calling it after
310  *      we had read the existing part of chain and partial points to the last
311  *      triple of that (one with zero ->key). Upon the exit we have the same
312  *      picture as after the successful ext4_get_block(), except that in one
313  *      place chain is disconnected - *branch->p is still zero (we did not
314  *      set the last link), but branch->key contains the number that should
315  *      be placed into *branch->p to fill that gap.
316  *
317  *      If allocation fails we free all blocks we've allocated (and forget
318  *      their buffer_heads) and return the error value the from failed
319  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
320  *      as described above and return 0.
321  */
322 static int ext4_alloc_branch(handle_t *handle,
323                              struct ext4_allocation_request *ar,
324                              int indirect_blks, ext4_lblk_t *offsets,
325                              Indirect *branch)
326 {
327         struct buffer_head *            bh;
328         ext4_fsblk_t                    b, new_blocks[4];
329         __le32                          *p;
330         int                             i, j, err, len = 1;
331
332         for (i = 0; i <= indirect_blks; i++) {
333                 if (i == indirect_blks) {
334                         new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
335                 } else
336                         ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
337                                         ar->inode, ar->goal,
338                                         ar->flags & EXT4_MB_DELALLOC_RESERVED,
339                                         NULL, &err);
340                 if (err) {
341                         i--;
342                         goto failed;
343                 }
344                 branch[i].key = cpu_to_le32(new_blocks[i]);
345                 if (i == 0)
346                         continue;
347
348                 bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
349                 if (unlikely(!bh)) {
350                         err = -ENOMEM;
351                         goto failed;
352                 }
353                 lock_buffer(bh);
354                 BUFFER_TRACE(bh, "call get_create_access");
355                 err = ext4_journal_get_create_access(handle, bh);
356                 if (err) {
357                         unlock_buffer(bh);
358                         goto failed;
359                 }
360
361                 memset(bh->b_data, 0, bh->b_size);
362                 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
363                 b = new_blocks[i];
364
365                 if (i == indirect_blks)
366                         len = ar->len;
367                 for (j = 0; j < len; j++)
368                         *p++ = cpu_to_le32(b++);
369
370                 BUFFER_TRACE(bh, "marking uptodate");
371                 set_buffer_uptodate(bh);
372                 unlock_buffer(bh);
373
374                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
375                 err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
376                 if (err)
377                         goto failed;
378         }
379         return 0;
380 failed:
381         for (; i >= 0; i--) {
382                 /*
383                  * We want to ext4_forget() only freshly allocated indirect
384                  * blocks.  Buffer for new_blocks[i-1] is at branch[i].bh and
385                  * buffer at branch[0].bh is indirect block / inode already
386                  * existing before ext4_alloc_branch() was called.
387                  */
388                 if (i > 0 && i != indirect_blks && branch[i].bh)
389                         ext4_forget(handle, 1, ar->inode, branch[i].bh,
390                                     branch[i].bh->b_blocknr);
391                 ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
392                                  (i == indirect_blks) ? ar->len : 1, 0);
393         }
394         return err;
395 }
396
397 /**
398  * ext4_splice_branch - splice the allocated branch onto inode.
399  * @handle: handle for this transaction
400  * @inode: owner
401  * @block: (logical) number of block we are adding
402  * @chain: chain of indirect blocks (with a missing link - see
403  *      ext4_alloc_branch)
404  * @where: location of missing link
405  * @num:   number of indirect blocks we are adding
406  * @blks:  number of direct blocks we are adding
407  *
408  * This function fills the missing link and does all housekeeping needed in
409  * inode (->i_blocks, etc.). In case of success we end up with the full
410  * chain to new block and return 0.
411  */
412 static int ext4_splice_branch(handle_t *handle,
413                               struct ext4_allocation_request *ar,
414                               Indirect *where, int num)
415 {
416         int i;
417         int err = 0;
418         ext4_fsblk_t current_block;
419
420         /*
421          * If we're splicing into a [td]indirect block (as opposed to the
422          * inode) then we need to get write access to the [td]indirect block
423          * before the splice.
424          */
425         if (where->bh) {
426                 BUFFER_TRACE(where->bh, "get_write_access");
427                 err = ext4_journal_get_write_access(handle, where->bh);
428                 if (err)
429                         goto err_out;
430         }
431         /* That's it */
432
433         *where->p = where->key;
434
435         /*
436          * Update the host buffer_head or inode to point to more just allocated
437          * direct blocks blocks
438          */
439         if (num == 0 && ar->len > 1) {
440                 current_block = le32_to_cpu(where->key) + 1;
441                 for (i = 1; i < ar->len; i++)
442                         *(where->p + i) = cpu_to_le32(current_block++);
443         }
444
445         /* We are done with atomic stuff, now do the rest of housekeeping */
446         /* had we spliced it onto indirect block? */
447         if (where->bh) {
448                 /*
449                  * If we spliced it onto an indirect block, we haven't
450                  * altered the inode.  Note however that if it is being spliced
451                  * onto an indirect block at the very end of the file (the
452                  * file is growing) then we *will* alter the inode to reflect
453                  * the new i_size.  But that is not done here - it is done in
454                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
455                  */
456                 jbd_debug(5, "splicing indirect only\n");
457                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
458                 err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
459                 if (err)
460                         goto err_out;
461         } else {
462                 /*
463                  * OK, we spliced it into the inode itself on a direct block.
464                  */
465                 ext4_mark_inode_dirty(handle, ar->inode);
466                 jbd_debug(5, "splicing direct\n");
467         }
468         return err;
469
470 err_out:
471         for (i = 1; i <= num; i++) {
472                 /*
473                  * branch[i].bh is newly allocated, so there is no
474                  * need to revoke the block, which is why we don't
475                  * need to set EXT4_FREE_BLOCKS_METADATA.
476                  */
477                 ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
478                                  EXT4_FREE_BLOCKS_FORGET);
479         }
480         ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
481                          ar->len, 0);
482
483         return err;
484 }
485
486 /*
487  * The ext4_ind_map_blocks() function handles non-extents inodes
488  * (i.e., using the traditional indirect/double-indirect i_blocks
489  * scheme) for ext4_map_blocks().
490  *
491  * Allocation strategy is simple: if we have to allocate something, we will
492  * have to go the whole way to leaf. So let's do it before attaching anything
493  * to tree, set linkage between the newborn blocks, write them if sync is
494  * required, recheck the path, free and repeat if check fails, otherwise
495  * set the last missing link (that will protect us from any truncate-generated
496  * removals - all blocks on the path are immune now) and possibly force the
497  * write on the parent block.
498  * That has a nice additional property: no special recovery from the failed
499  * allocations is needed - we simply release blocks and do not touch anything
500  * reachable from inode.
501  *
502  * `handle' can be NULL if create == 0.
503  *
504  * return > 0, # of blocks mapped or allocated.
505  * return = 0, if plain lookup failed.
506  * return < 0, error case.
507  *
508  * The ext4_ind_get_blocks() function should be called with
509  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
510  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
511  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
512  * blocks.
513  */
514 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
515                         struct ext4_map_blocks *map,
516                         int flags)
517 {
518         struct ext4_allocation_request ar;
519         int err = -EIO;
520         ext4_lblk_t offsets[4];
521         Indirect chain[4];
522         Indirect *partial;
523         int indirect_blks;
524         int blocks_to_boundary = 0;
525         int depth;
526         int count = 0;
527         ext4_fsblk_t first_block = 0;
528
529         trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
530         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
531         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
532         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
533                                    &blocks_to_boundary);
534
535         if (depth == 0)
536                 goto out;
537
538         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
539
540         /* Simplest case - block found, no allocation needed */
541         if (!partial) {
542                 first_block = le32_to_cpu(chain[depth - 1].key);
543                 count++;
544                 /*map more blocks*/
545                 while (count < map->m_len && count <= blocks_to_boundary) {
546                         ext4_fsblk_t blk;
547
548                         blk = le32_to_cpu(*(chain[depth-1].p + count));
549
550                         if (blk == first_block + count)
551                                 count++;
552                         else
553                                 break;
554                 }
555                 goto got_it;
556         }
557
558         /* Next simple case - plain lookup failed */
559         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
560                 unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
561                 int i;
562
563                 /*
564                  * Count number blocks in a subtree under 'partial'. At each
565                  * level we count number of complete empty subtrees beyond
566                  * current offset and then descend into the subtree only
567                  * partially beyond current offset.
568                  */
569                 count = 0;
570                 for (i = partial - chain + 1; i < depth; i++)
571                         count = count * epb + (epb - offsets[i] - 1);
572                 count++;
573                 /* Fill in size of a hole we found */
574                 map->m_pblk = 0;
575                 map->m_len = min_t(unsigned int, map->m_len, count);
576                 goto cleanup;
577         }
578
579         /* Failed read of indirect block */
580         if (err == -EIO)
581                 goto cleanup;
582
583         /*
584          * Okay, we need to do block allocation.
585         */
586         if (ext4_has_feature_bigalloc(inode->i_sb)) {
587                 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
588                                  "non-extent mapped inodes with bigalloc");
589                 return -EFSCORRUPTED;
590         }
591
592         /* Set up for the direct block allocation */
593         memset(&ar, 0, sizeof(ar));
594         ar.inode = inode;
595         ar.logical = map->m_lblk;
596         if (S_ISREG(inode->i_mode))
597                 ar.flags = EXT4_MB_HINT_DATA;
598         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
599                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
600         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
601                 ar.flags |= EXT4_MB_USE_RESERVED;
602
603         ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
604
605         /* the number of blocks need to allocate for [d,t]indirect blocks */
606         indirect_blks = (chain + depth) - partial - 1;
607
608         /*
609          * Next look up the indirect map to count the totoal number of
610          * direct blocks to allocate for this branch.
611          */
612         ar.len = ext4_blks_to_allocate(partial, indirect_blks,
613                                        map->m_len, blocks_to_boundary);
614
615         /*
616          * Block out ext4_truncate while we alter the tree
617          */
618         err = ext4_alloc_branch(handle, &ar, indirect_blks,
619                                 offsets + (partial - chain), partial);
620
621         /*
622          * The ext4_splice_branch call will free and forget any buffers
623          * on the new chain if there is a failure, but that risks using
624          * up transaction credits, especially for bitmaps where the
625          * credits cannot be returned.  Can we handle this somehow?  We
626          * may need to return -EAGAIN upwards in the worst case.  --sct
627          */
628         if (!err)
629                 err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
630         if (err)
631                 goto cleanup;
632
633         map->m_flags |= EXT4_MAP_NEW;
634
635         ext4_update_inode_fsync_trans(handle, inode, 1);
636         count = ar.len;
637 got_it:
638         map->m_flags |= EXT4_MAP_MAPPED;
639         map->m_pblk = le32_to_cpu(chain[depth-1].key);
640         map->m_len = count;
641         if (count > blocks_to_boundary)
642                 map->m_flags |= EXT4_MAP_BOUNDARY;
643         err = count;
644         /* Clean up and exit */
645         partial = chain + depth - 1;    /* the whole chain */
646 cleanup:
647         while (partial > chain) {
648                 BUFFER_TRACE(partial->bh, "call brelse");
649                 brelse(partial->bh);
650                 partial--;
651         }
652 out:
653         trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
654         return err;
655 }
656
657 /*
658  * Calculate the number of metadata blocks need to reserve
659  * to allocate a new block at @lblocks for non extent file based file
660  */
661 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
662 {
663         struct ext4_inode_info *ei = EXT4_I(inode);
664         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
665         int blk_bits;
666
667         if (lblock < EXT4_NDIR_BLOCKS)
668                 return 0;
669
670         lblock -= EXT4_NDIR_BLOCKS;
671
672         if (ei->i_da_metadata_calc_len &&
673             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
674                 ei->i_da_metadata_calc_len++;
675                 return 0;
676         }
677         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
678         ei->i_da_metadata_calc_len = 1;
679         blk_bits = order_base_2(lblock);
680         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
681 }
682
683 /*
684  * Calculate number of indirect blocks touched by mapping @nrblocks logically
685  * contiguous blocks
686  */
687 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
688 {
689         /*
690          * With N contiguous data blocks, we need at most
691          * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
692          * 2 dindirect blocks, and 1 tindirect block
693          */
694         return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
695 }
696
697 /*
698  * Truncate transactions can be complex and absolutely huge.  So we need to
699  * be able to restart the transaction at a conventient checkpoint to make
700  * sure we don't overflow the journal.
701  *
702  * Try to extend this transaction for the purposes of truncation.  If
703  * extend fails, we need to propagate the failure up and restart the
704  * transaction in the top-level truncate loop. --sct
705  *
706  * Returns 0 if we managed to create more room.  If we can't create more
707  * room, and the transaction must be restarted we return 1.
708  */
709 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
710 {
711         if (!ext4_handle_valid(handle))
712                 return 0;
713         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
714                 return 0;
715         if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
716                 return 0;
717         return 1;
718 }
719
720 /*
721  * Probably it should be a library function... search for first non-zero word
722  * or memcmp with zero_page, whatever is better for particular architecture.
723  * Linus?
724  */
725 static inline int all_zeroes(__le32 *p, __le32 *q)
726 {
727         while (p < q)
728                 if (*p++)
729                         return 0;
730         return 1;
731 }
732
733 /**
734  *      ext4_find_shared - find the indirect blocks for partial truncation.
735  *      @inode:   inode in question
736  *      @depth:   depth of the affected branch
737  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
738  *      @chain:   place to store the pointers to partial indirect blocks
739  *      @top:     place to the (detached) top of branch
740  *
741  *      This is a helper function used by ext4_truncate().
742  *
743  *      When we do truncate() we may have to clean the ends of several
744  *      indirect blocks but leave the blocks themselves alive. Block is
745  *      partially truncated if some data below the new i_size is referred
746  *      from it (and it is on the path to the first completely truncated
747  *      data block, indeed).  We have to free the top of that path along
748  *      with everything to the right of the path. Since no allocation
749  *      past the truncation point is possible until ext4_truncate()
750  *      finishes, we may safely do the latter, but top of branch may
751  *      require special attention - pageout below the truncation point
752  *      might try to populate it.
753  *
754  *      We atomically detach the top of branch from the tree, store the
755  *      block number of its root in *@top, pointers to buffer_heads of
756  *      partially truncated blocks - in @chain[].bh and pointers to
757  *      their last elements that should not be removed - in
758  *      @chain[].p. Return value is the pointer to last filled element
759  *      of @chain.
760  *
761  *      The work left to caller to do the actual freeing of subtrees:
762  *              a) free the subtree starting from *@top
763  *              b) free the subtrees whose roots are stored in
764  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
765  *              c) free the subtrees growing from the inode past the @chain[0].
766  *                      (no partially truncated stuff there).  */
767
768 static Indirect *ext4_find_shared(struct inode *inode, int depth,
769                                   ext4_lblk_t offsets[4], Indirect chain[4],
770                                   __le32 *top)
771 {
772         Indirect *partial, *p;
773         int k, err;
774
775         *top = 0;
776         /* Make k index the deepest non-null offset + 1 */
777         for (k = depth; k > 1 && !offsets[k-1]; k--)
778                 ;
779         partial = ext4_get_branch(inode, k, offsets, chain, &err);
780         /* Writer: pointers */
781         if (!partial)
782                 partial = chain + k-1;
783         /*
784          * If the branch acquired continuation since we've looked at it -
785          * fine, it should all survive and (new) top doesn't belong to us.
786          */
787         if (!partial->key && *partial->p)
788                 /* Writer: end */
789                 goto no_top;
790         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
791                 ;
792         /*
793          * OK, we've found the last block that must survive. The rest of our
794          * branch should be detached before unlocking. However, if that rest
795          * of branch is all ours and does not grow immediately from the inode
796          * it's easier to cheat and just decrement partial->p.
797          */
798         if (p == chain + k - 1 && p > chain) {
799                 p->p--;
800         } else {
801                 *top = *p->p;
802                 /* Nope, don't do this in ext4.  Must leave the tree intact */
803 #if 0
804                 *p->p = 0;
805 #endif
806         }
807         /* Writer: end */
808
809         while (partial > p) {
810                 brelse(partial->bh);
811                 partial--;
812         }
813 no_top:
814         return partial;
815 }
816
817 /*
818  * Zero a number of block pointers in either an inode or an indirect block.
819  * If we restart the transaction we must again get write access to the
820  * indirect block for further modification.
821  *
822  * We release `count' blocks on disk, but (last - first) may be greater
823  * than `count' because there can be holes in there.
824  *
825  * Return 0 on success, 1 on invalid block range
826  * and < 0 on fatal error.
827  */
828 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
829                              struct buffer_head *bh,
830                              ext4_fsblk_t block_to_free,
831                              unsigned long count, __le32 *first,
832                              __le32 *last)
833 {
834         __le32 *p;
835         int     flags = EXT4_FREE_BLOCKS_VALIDATED;
836         int     err;
837
838         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
839                 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
840         else if (ext4_should_journal_data(inode))
841                 flags |= EXT4_FREE_BLOCKS_FORGET;
842
843         if (!ext4_inode_block_valid(inode, block_to_free, count)) {
844                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
845                                  "blocks %llu len %lu",
846                                  (unsigned long long) block_to_free, count);
847                 return 1;
848         }
849
850         if (try_to_extend_transaction(handle, inode)) {
851                 if (bh) {
852                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
853                         err = ext4_handle_dirty_metadata(handle, inode, bh);
854                         if (unlikely(err))
855                                 goto out_err;
856                 }
857                 err = ext4_mark_inode_dirty(handle, inode);
858                 if (unlikely(err))
859                         goto out_err;
860                 err = ext4_truncate_restart_trans(handle, inode,
861                                         ext4_blocks_for_truncate(inode));
862                 if (unlikely(err))
863                         goto out_err;
864                 if (bh) {
865                         BUFFER_TRACE(bh, "retaking write access");
866                         err = ext4_journal_get_write_access(handle, bh);
867                         if (unlikely(err))
868                                 goto out_err;
869                 }
870         }
871
872         for (p = first; p < last; p++)
873                 *p = 0;
874
875         ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
876         return 0;
877 out_err:
878         ext4_std_error(inode->i_sb, err);
879         return err;
880 }
881
882 /**
883  * ext4_free_data - free a list of data blocks
884  * @handle:     handle for this transaction
885  * @inode:      inode we are dealing with
886  * @this_bh:    indirect buffer_head which contains *@first and *@last
887  * @first:      array of block numbers
888  * @last:       points immediately past the end of array
889  *
890  * We are freeing all blocks referred from that array (numbers are stored as
891  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
892  *
893  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
894  * blocks are contiguous then releasing them at one time will only affect one
895  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
896  * actually use a lot of journal space.
897  *
898  * @this_bh will be %NULL if @first and @last point into the inode's direct
899  * block pointers.
900  */
901 static void ext4_free_data(handle_t *handle, struct inode *inode,
902                            struct buffer_head *this_bh,
903                            __le32 *first, __le32 *last)
904 {
905         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
906         unsigned long count = 0;            /* Number of blocks in the run */
907         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
908                                                corresponding to
909                                                block_to_free */
910         ext4_fsblk_t nr;                    /* Current block # */
911         __le32 *p;                          /* Pointer into inode/ind
912                                                for current block */
913         int err = 0;
914
915         if (this_bh) {                          /* For indirect block */
916                 BUFFER_TRACE(this_bh, "get_write_access");
917                 err = ext4_journal_get_write_access(handle, this_bh);
918                 /* Important: if we can't update the indirect pointers
919                  * to the blocks, we can't free them. */
920                 if (err)
921                         return;
922         }
923
924         for (p = first; p < last; p++) {
925                 nr = le32_to_cpu(*p);
926                 if (nr) {
927                         /* accumulate blocks to free if they're contiguous */
928                         if (count == 0) {
929                                 block_to_free = nr;
930                                 block_to_free_p = p;
931                                 count = 1;
932                         } else if (nr == block_to_free + count) {
933                                 count++;
934                         } else {
935                                 err = ext4_clear_blocks(handle, inode, this_bh,
936                                                         block_to_free, count,
937                                                         block_to_free_p, p);
938                                 if (err)
939                                         break;
940                                 block_to_free = nr;
941                                 block_to_free_p = p;
942                                 count = 1;
943                         }
944                 }
945         }
946
947         if (!err && count > 0)
948                 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
949                                         count, block_to_free_p, p);
950         if (err < 0)
951                 /* fatal error */
952                 return;
953
954         if (this_bh) {
955                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
956
957                 /*
958                  * The buffer head should have an attached journal head at this
959                  * point. However, if the data is corrupted and an indirect
960                  * block pointed to itself, it would have been detached when
961                  * the block was cleared. Check for this instead of OOPSing.
962                  */
963                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
964                         ext4_handle_dirty_metadata(handle, inode, this_bh);
965                 else
966                         EXT4_ERROR_INODE(inode,
967                                          "circular indirect block detected at "
968                                          "block %llu",
969                                 (unsigned long long) this_bh->b_blocknr);
970         }
971 }
972
973 /**
974  *      ext4_free_branches - free an array of branches
975  *      @handle: JBD handle for this transaction
976  *      @inode: inode we are dealing with
977  *      @parent_bh: the buffer_head which contains *@first and *@last
978  *      @first: array of block numbers
979  *      @last:  pointer immediately past the end of array
980  *      @depth: depth of the branches to free
981  *
982  *      We are freeing all blocks referred from these branches (numbers are
983  *      stored as little-endian 32-bit) and updating @inode->i_blocks
984  *      appropriately.
985  */
986 static void ext4_free_branches(handle_t *handle, struct inode *inode,
987                                struct buffer_head *parent_bh,
988                                __le32 *first, __le32 *last, int depth)
989 {
990         ext4_fsblk_t nr;
991         __le32 *p;
992
993         if (ext4_handle_is_aborted(handle))
994                 return;
995
996         if (depth--) {
997                 struct buffer_head *bh;
998                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
999                 p = last;
1000                 while (--p >= first) {
1001                         nr = le32_to_cpu(*p);
1002                         if (!nr)
1003                                 continue;               /* A hole */
1004
1005                         if (!ext4_inode_block_valid(inode, nr, 1)) {
1006                                 EXT4_ERROR_INODE(inode,
1007                                                  "invalid indirect mapped "
1008                                                  "block %lu (level %d)",
1009                                                  (unsigned long) nr, depth);
1010                                 break;
1011                         }
1012
1013                         /* Go read the buffer for the next level down */
1014                         bh = sb_bread(inode->i_sb, nr);
1015
1016                         /*
1017                          * A read failure? Report error and clear slot
1018                          * (should be rare).
1019                          */
1020                         if (!bh) {
1021                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
1022                                                        "Read failure");
1023                                 continue;
1024                         }
1025
1026                         /* This zaps the entire block.  Bottom up. */
1027                         BUFFER_TRACE(bh, "free child branches");
1028                         ext4_free_branches(handle, inode, bh,
1029                                         (__le32 *) bh->b_data,
1030                                         (__le32 *) bh->b_data + addr_per_block,
1031                                         depth);
1032                         brelse(bh);
1033
1034                         /*
1035                          * Everything below this this pointer has been
1036                          * released.  Now let this top-of-subtree go.
1037                          *
1038                          * We want the freeing of this indirect block to be
1039                          * atomic in the journal with the updating of the
1040                          * bitmap block which owns it.  So make some room in
1041                          * the journal.
1042                          *
1043                          * We zero the parent pointer *after* freeing its
1044                          * pointee in the bitmaps, so if extend_transaction()
1045                          * for some reason fails to put the bitmap changes and
1046                          * the release into the same transaction, recovery
1047                          * will merely complain about releasing a free block,
1048                          * rather than leaking blocks.
1049                          */
1050                         if (ext4_handle_is_aborted(handle))
1051                                 return;
1052                         if (try_to_extend_transaction(handle, inode)) {
1053                                 ext4_mark_inode_dirty(handle, inode);
1054                                 ext4_truncate_restart_trans(handle, inode,
1055                                             ext4_blocks_for_truncate(inode));
1056                         }
1057
1058                         /*
1059                          * The forget flag here is critical because if
1060                          * we are journaling (and not doing data
1061                          * journaling), we have to make sure a revoke
1062                          * record is written to prevent the journal
1063                          * replay from overwriting the (former)
1064                          * indirect block if it gets reallocated as a
1065                          * data block.  This must happen in the same
1066                          * transaction where the data blocks are
1067                          * actually freed.
1068                          */
1069                         ext4_free_blocks(handle, inode, NULL, nr, 1,
1070                                          EXT4_FREE_BLOCKS_METADATA|
1071                                          EXT4_FREE_BLOCKS_FORGET);
1072
1073                         if (parent_bh) {
1074                                 /*
1075                                  * The block which we have just freed is
1076                                  * pointed to by an indirect block: journal it
1077                                  */
1078                                 BUFFER_TRACE(parent_bh, "get_write_access");
1079                                 if (!ext4_journal_get_write_access(handle,
1080                                                                    parent_bh)){
1081                                         *p = 0;
1082                                         BUFFER_TRACE(parent_bh,
1083                                         "call ext4_handle_dirty_metadata");
1084                                         ext4_handle_dirty_metadata(handle,
1085                                                                    inode,
1086                                                                    parent_bh);
1087                                 }
1088                         }
1089                 }
1090         } else {
1091                 /* We have reached the bottom of the tree. */
1092                 BUFFER_TRACE(parent_bh, "free data blocks");
1093                 ext4_free_data(handle, inode, parent_bh, first, last);
1094         }
1095 }
1096
1097 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1098 {
1099         struct ext4_inode_info *ei = EXT4_I(inode);
1100         __le32 *i_data = ei->i_data;
1101         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1102         ext4_lblk_t offsets[4];
1103         Indirect chain[4];
1104         Indirect *partial;
1105         __le32 nr = 0;
1106         int n = 0;
1107         ext4_lblk_t last_block, max_block;
1108         unsigned blocksize = inode->i_sb->s_blocksize;
1109
1110         last_block = (inode->i_size + blocksize-1)
1111                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1112         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1113                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1114
1115         if (last_block != max_block) {
1116                 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1117                 if (n == 0)
1118                         return;
1119         }
1120
1121         ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1122
1123         /*
1124          * The orphan list entry will now protect us from any crash which
1125          * occurs before the truncate completes, so it is now safe to propagate
1126          * the new, shorter inode size (held for now in i_size) into the
1127          * on-disk inode. We do this via i_disksize, which is the value which
1128          * ext4 *really* writes onto the disk inode.
1129          */
1130         ei->i_disksize = inode->i_size;
1131
1132         if (last_block == max_block) {
1133                 /*
1134                  * It is unnecessary to free any data blocks if last_block is
1135                  * equal to the indirect block limit.
1136                  */
1137                 return;
1138         } else if (n == 1) {            /* direct blocks */
1139                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1140                                i_data + EXT4_NDIR_BLOCKS);
1141                 goto do_indirects;
1142         }
1143
1144         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1145         /* Kill the top of shared branch (not detached) */
1146         if (nr) {
1147                 if (partial == chain) {
1148                         /* Shared branch grows from the inode */
1149                         ext4_free_branches(handle, inode, NULL,
1150                                            &nr, &nr+1, (chain+n-1) - partial);
1151                         *partial->p = 0;
1152                         /*
1153                          * We mark the inode dirty prior to restart,
1154                          * and prior to stop.  No need for it here.
1155                          */
1156                 } else {
1157                         /* Shared branch grows from an indirect block */
1158                         BUFFER_TRACE(partial->bh, "get_write_access");
1159                         ext4_free_branches(handle, inode, partial->bh,
1160                                         partial->p,
1161                                         partial->p+1, (chain+n-1) - partial);
1162                 }
1163         }
1164         /* Clear the ends of indirect blocks on the shared branch */
1165         while (partial > chain) {
1166                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1167                                    (__le32*)partial->bh->b_data+addr_per_block,
1168                                    (chain+n-1) - partial);
1169                 BUFFER_TRACE(partial->bh, "call brelse");
1170                 brelse(partial->bh);
1171                 partial--;
1172         }
1173 do_indirects:
1174         /* Kill the remaining (whole) subtrees */
1175         switch (offsets[0]) {
1176         default:
1177                 nr = i_data[EXT4_IND_BLOCK];
1178                 if (nr) {
1179                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1180                         i_data[EXT4_IND_BLOCK] = 0;
1181                 }
1182         case EXT4_IND_BLOCK:
1183                 nr = i_data[EXT4_DIND_BLOCK];
1184                 if (nr) {
1185                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1186                         i_data[EXT4_DIND_BLOCK] = 0;
1187                 }
1188         case EXT4_DIND_BLOCK:
1189                 nr = i_data[EXT4_TIND_BLOCK];
1190                 if (nr) {
1191                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1192                         i_data[EXT4_TIND_BLOCK] = 0;
1193                 }
1194         case EXT4_TIND_BLOCK:
1195                 ;
1196         }
1197 }
1198
1199 /**
1200  *      ext4_ind_remove_space - remove space from the range
1201  *      @handle: JBD handle for this transaction
1202  *      @inode: inode we are dealing with
1203  *      @start: First block to remove
1204  *      @end:   One block after the last block to remove (exclusive)
1205  *
1206  *      Free the blocks in the defined range (end is exclusive endpoint of
1207  *      range). This is used by ext4_punch_hole().
1208  */
1209 int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1210                           ext4_lblk_t start, ext4_lblk_t end)
1211 {
1212         struct ext4_inode_info *ei = EXT4_I(inode);
1213         __le32 *i_data = ei->i_data;
1214         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1215         ext4_lblk_t offsets[4], offsets2[4];
1216         Indirect chain[4], chain2[4];
1217         Indirect *partial, *partial2;
1218         Indirect *p = NULL, *p2 = NULL;
1219         ext4_lblk_t max_block;
1220         __le32 nr = 0, nr2 = 0;
1221         int n = 0, n2 = 0;
1222         unsigned blocksize = inode->i_sb->s_blocksize;
1223
1224         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1225                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1226         if (end >= max_block)
1227                 end = max_block;
1228         if ((start >= end) || (start > max_block))
1229                 return 0;
1230
1231         n = ext4_block_to_path(inode, start, offsets, NULL);
1232         n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1233
1234         BUG_ON(n > n2);
1235
1236         if ((n == 1) && (n == n2)) {
1237                 /* We're punching only within direct block range */
1238                 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1239                                i_data + offsets2[0]);
1240                 return 0;
1241         } else if (n2 > n) {
1242                 /*
1243                  * Start and end are on a different levels so we're going to
1244                  * free partial block at start, and partial block at end of
1245                  * the range. If there are some levels in between then
1246                  * do_indirects label will take care of that.
1247                  */
1248
1249                 if (n == 1) {
1250                         /*
1251                          * Start is at the direct block level, free
1252                          * everything to the end of the level.
1253                          */
1254                         ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1255                                        i_data + EXT4_NDIR_BLOCKS);
1256                         goto end_range;
1257                 }
1258
1259
1260                 partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1261                 if (nr) {
1262                         if (partial == chain) {
1263                                 /* Shared branch grows from the inode */
1264                                 ext4_free_branches(handle, inode, NULL,
1265                                            &nr, &nr+1, (chain+n-1) - partial);
1266                                 *partial->p = 0;
1267                         } else {
1268                                 /* Shared branch grows from an indirect block */
1269                                 BUFFER_TRACE(partial->bh, "get_write_access");
1270                                 ext4_free_branches(handle, inode, partial->bh,
1271                                         partial->p,
1272                                         partial->p+1, (chain+n-1) - partial);
1273                         }
1274                 }
1275
1276                 /*
1277                  * Clear the ends of indirect blocks on the shared branch
1278                  * at the start of the range
1279                  */
1280                 while (partial > chain) {
1281                         ext4_free_branches(handle, inode, partial->bh,
1282                                 partial->p + 1,
1283                                 (__le32 *)partial->bh->b_data+addr_per_block,
1284                                 (chain+n-1) - partial);
1285                         partial--;
1286                 }
1287
1288 end_range:
1289                 partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1290                 if (nr2) {
1291                         if (partial2 == chain2) {
1292                                 /*
1293                                  * Remember, end is exclusive so here we're at
1294                                  * the start of the next level we're not going
1295                                  * to free. Everything was covered by the start
1296                                  * of the range.
1297                                  */
1298                                 goto do_indirects;
1299                         }
1300                 } else {
1301                         /*
1302                          * ext4_find_shared returns Indirect structure which
1303                          * points to the last element which should not be
1304                          * removed by truncate. But this is end of the range
1305                          * in punch_hole so we need to point to the next element
1306                          */
1307                         partial2->p++;
1308                 }
1309
1310                 /*
1311                  * Clear the ends of indirect blocks on the shared branch
1312                  * at the end of the range
1313                  */
1314                 while (partial2 > chain2) {
1315                         ext4_free_branches(handle, inode, partial2->bh,
1316                                            (__le32 *)partial2->bh->b_data,
1317                                            partial2->p,
1318                                            (chain2+n2-1) - partial2);
1319                         partial2--;
1320                 }
1321                 goto do_indirects;
1322         }
1323
1324         /* Punch happened within the same level (n == n2) */
1325         partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1326         partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1327
1328         /* Free top, but only if partial2 isn't its subtree. */
1329         if (nr) {
1330                 int level = min(partial - chain, partial2 - chain2);
1331                 int i;
1332                 int subtree = 1;
1333
1334                 for (i = 0; i <= level; i++) {
1335                         if (offsets[i] != offsets2[i]) {
1336                                 subtree = 0;
1337                                 break;
1338                         }
1339                 }
1340
1341                 if (!subtree) {
1342                         if (partial == chain) {
1343                                 /* Shared branch grows from the inode */
1344                                 ext4_free_branches(handle, inode, NULL,
1345                                                    &nr, &nr+1,
1346                                                    (chain+n-1) - partial);
1347                                 *partial->p = 0;
1348                         } else {
1349                                 /* Shared branch grows from an indirect block */
1350                                 BUFFER_TRACE(partial->bh, "get_write_access");
1351                                 ext4_free_branches(handle, inode, partial->bh,
1352                                                    partial->p,
1353                                                    partial->p+1,
1354                                                    (chain+n-1) - partial);
1355                         }
1356                 }
1357         }
1358
1359         if (!nr2) {
1360                 /*
1361                  * ext4_find_shared returns Indirect structure which
1362                  * points to the last element which should not be
1363                  * removed by truncate. But this is end of the range
1364                  * in punch_hole so we need to point to the next element
1365                  */
1366                 partial2->p++;
1367         }
1368
1369         while (partial > chain || partial2 > chain2) {
1370                 int depth = (chain+n-1) - partial;
1371                 int depth2 = (chain2+n2-1) - partial2;
1372
1373                 if (partial > chain && partial2 > chain2 &&
1374                     partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1375                         /*
1376                          * We've converged on the same block. Clear the range,
1377                          * then we're done.
1378                          */
1379                         ext4_free_branches(handle, inode, partial->bh,
1380                                            partial->p + 1,
1381                                            partial2->p,
1382                                            (chain+n-1) - partial);
1383                         goto cleanup;
1384                 }
1385
1386                 /*
1387                  * The start and end partial branches may not be at the same
1388                  * level even though the punch happened within one level. So, we
1389                  * give them a chance to arrive at the same level, then walk
1390                  * them in step with each other until we converge on the same
1391                  * block.
1392                  */
1393                 if (partial > chain && depth <= depth2) {
1394                         ext4_free_branches(handle, inode, partial->bh,
1395                                            partial->p + 1,
1396                                            (__le32 *)partial->bh->b_data+addr_per_block,
1397                                            (chain+n-1) - partial);
1398                         partial--;
1399                 }
1400                 if (partial2 > chain2 && depth2 <= depth) {
1401                         ext4_free_branches(handle, inode, partial2->bh,
1402                                            (__le32 *)partial2->bh->b_data,
1403                                            partial2->p,
1404                                            (chain2+n2-1) - partial2);
1405                         partial2--;
1406                 }
1407         }
1408
1409 cleanup:
1410         while (p && p > chain) {
1411                 BUFFER_TRACE(p->bh, "call brelse");
1412                 brelse(p->bh);
1413                 p--;
1414         }
1415         while (p2 && p2 > chain2) {
1416                 BUFFER_TRACE(p2->bh, "call brelse");
1417                 brelse(p2->bh);
1418                 p2--;
1419         }
1420         return 0;
1421
1422 do_indirects:
1423         /* Kill the remaining (whole) subtrees */
1424         switch (offsets[0]) {
1425         default:
1426                 if (++n >= n2)
1427                         break;
1428                 nr = i_data[EXT4_IND_BLOCK];
1429                 if (nr) {
1430                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1431                         i_data[EXT4_IND_BLOCK] = 0;
1432                 }
1433         case EXT4_IND_BLOCK:
1434                 if (++n >= n2)
1435                         break;
1436                 nr = i_data[EXT4_DIND_BLOCK];
1437                 if (nr) {
1438                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1439                         i_data[EXT4_DIND_BLOCK] = 0;
1440                 }
1441         case EXT4_DIND_BLOCK:
1442                 if (++n >= n2)
1443                         break;
1444                 nr = i_data[EXT4_TIND_BLOCK];
1445                 if (nr) {
1446                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1447                         i_data[EXT4_TIND_BLOCK] = 0;
1448                 }
1449         case EXT4_TIND_BLOCK:
1450                 ;
1451         }
1452         goto cleanup;
1453 }