GNU Linux-libre 5.4.257-gnu1
[releases.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         unsigned int key;
152         int ret = -EIO;
153
154         *err = 0;
155         /* i_data is not going away, no lock needed */
156         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
157         if (!p->key)
158                 goto no_block;
159         while (--depth) {
160                 key = le32_to_cpu(p->key);
161                 if (key > ext4_blocks_count(EXT4_SB(sb)->s_es)) {
162                         /* the block was out of range */
163                         ret = -EFSCORRUPTED;
164                         goto failure;
165                 }
166                 bh = sb_getblk(sb, key);
167                 if (unlikely(!bh)) {
168                         ret = -ENOMEM;
169                         goto failure;
170                 }
171
172                 if (!bh_uptodate_or_lock(bh)) {
173                         if (bh_submit_read(bh) < 0) {
174                                 put_bh(bh);
175                                 goto failure;
176                         }
177                         /* validate block references */
178                         if (ext4_check_indirect_blockref(inode, bh)) {
179                                 put_bh(bh);
180                                 goto failure;
181                         }
182                 }
183
184                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
185                 /* Reader: end */
186                 if (!p->key)
187                         goto no_block;
188         }
189         return NULL;
190
191 failure:
192         *err = ret;
193 no_block:
194         return p;
195 }
196
197 /**
198  *      ext4_find_near - find a place for allocation with sufficient locality
199  *      @inode: owner
200  *      @ind: descriptor of indirect block.
201  *
202  *      This function returns the preferred place for block allocation.
203  *      It is used when heuristic for sequential allocation fails.
204  *      Rules are:
205  *        + if there is a block to the left of our position - allocate near it.
206  *        + if pointer will live in indirect block - allocate near that block.
207  *        + if pointer will live in inode - allocate in the same
208  *          cylinder group.
209  *
210  * In the latter case we colour the starting block by the callers PID to
211  * prevent it from clashing with concurrent allocations for a different inode
212  * in the same block group.   The PID is used here so that functionally related
213  * files will be close-by on-disk.
214  *
215  *      Caller must make sure that @ind is valid and will stay that way.
216  */
217 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
218 {
219         struct ext4_inode_info *ei = EXT4_I(inode);
220         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
221         __le32 *p;
222
223         /* Try to find previous block */
224         for (p = ind->p - 1; p >= start; p--) {
225                 if (*p)
226                         return le32_to_cpu(*p);
227         }
228
229         /* No such thing, so let's try location of indirect block */
230         if (ind->bh)
231                 return ind->bh->b_blocknr;
232
233         /*
234          * It is going to be referred to from the inode itself? OK, just put it
235          * into the same cylinder group then.
236          */
237         return ext4_inode_to_goal_block(inode);
238 }
239
240 /**
241  *      ext4_find_goal - find a preferred place for allocation.
242  *      @inode: owner
243  *      @block:  block we want
244  *      @partial: pointer to the last triple within a chain
245  *
246  *      Normally this function find the preferred place for block allocation,
247  *      returns it.
248  *      Because this is only used for non-extent files, we limit the block nr
249  *      to 32 bits.
250  */
251 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
252                                    Indirect *partial)
253 {
254         ext4_fsblk_t goal;
255
256         /*
257          * XXX need to get goal block from mballoc's data structures
258          */
259
260         goal = ext4_find_near(inode, partial);
261         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
262         return goal;
263 }
264
265 /**
266  *      ext4_blks_to_allocate - Look up the block map and count the number
267  *      of direct blocks need to be allocated for the given branch.
268  *
269  *      @branch: chain of indirect blocks
270  *      @k: number of blocks need for indirect blocks
271  *      @blks: number of data blocks to be mapped.
272  *      @blocks_to_boundary:  the offset in the indirect block
273  *
274  *      return the total number of blocks to be allocate, including the
275  *      direct and indirect blocks.
276  */
277 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
278                                  int blocks_to_boundary)
279 {
280         unsigned int count = 0;
281
282         /*
283          * Simple case, [t,d]Indirect block(s) has not allocated yet
284          * then it's clear blocks on that path have not allocated
285          */
286         if (k > 0) {
287                 /* right now we don't handle cross boundary allocation */
288                 if (blks < blocks_to_boundary + 1)
289                         count += blks;
290                 else
291                         count += blocks_to_boundary + 1;
292                 return count;
293         }
294
295         count++;
296         while (count < blks && count <= blocks_to_boundary &&
297                 le32_to_cpu(*(branch[0].p + count)) == 0) {
298                 count++;
299         }
300         return count;
301 }
302
303 /**
304  * ext4_alloc_branch() - allocate and set up a chain of blocks
305  * @handle: handle for this transaction
306  * @ar: structure describing the allocation request
307  * @indirect_blks: number of allocated indirect blocks
308  * @offsets: offsets (in the blocks) to store the pointers to next.
309  * @branch: place to store the chain in.
310  *
311  *      This function allocates blocks, zeroes out all but the last one,
312  *      links them into chain and (if we are synchronous) writes them to disk.
313  *      In other words, it prepares a branch that can be spliced onto the
314  *      inode. It stores the information about that chain in the branch[], in
315  *      the same format as ext4_get_branch() would do. We are calling it after
316  *      we had read the existing part of chain and partial points to the last
317  *      triple of that (one with zero ->key). Upon the exit we have the same
318  *      picture as after the successful ext4_get_block(), except that in one
319  *      place chain is disconnected - *branch->p is still zero (we did not
320  *      set the last link), but branch->key contains the number that should
321  *      be placed into *branch->p to fill that gap.
322  *
323  *      If allocation fails we free all blocks we've allocated (and forget
324  *      their buffer_heads) and return the error value the from failed
325  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
326  *      as described above and return 0.
327  */
328 static int ext4_alloc_branch(handle_t *handle,
329                              struct ext4_allocation_request *ar,
330                              int indirect_blks, ext4_lblk_t *offsets,
331                              Indirect *branch)
332 {
333         struct buffer_head *            bh;
334         ext4_fsblk_t                    b, new_blocks[4];
335         __le32                          *p;
336         int                             i, j, err, len = 1;
337
338         for (i = 0; i <= indirect_blks; i++) {
339                 if (i == indirect_blks) {
340                         new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
341                 } else
342                         ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
343                                         ar->inode, ar->goal,
344                                         ar->flags & EXT4_MB_DELALLOC_RESERVED,
345                                         NULL, &err);
346                 if (err) {
347                         i--;
348                         goto failed;
349                 }
350                 branch[i].key = cpu_to_le32(new_blocks[i]);
351                 if (i == 0)
352                         continue;
353
354                 bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
355                 if (unlikely(!bh)) {
356                         err = -ENOMEM;
357                         goto failed;
358                 }
359                 lock_buffer(bh);
360                 BUFFER_TRACE(bh, "call get_create_access");
361                 err = ext4_journal_get_create_access(handle, bh);
362                 if (err) {
363                         unlock_buffer(bh);
364                         goto failed;
365                 }
366
367                 memset(bh->b_data, 0, bh->b_size);
368                 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
369                 b = new_blocks[i];
370
371                 if (i == indirect_blks)
372                         len = ar->len;
373                 for (j = 0; j < len; j++)
374                         *p++ = cpu_to_le32(b++);
375
376                 BUFFER_TRACE(bh, "marking uptodate");
377                 set_buffer_uptodate(bh);
378                 unlock_buffer(bh);
379
380                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
381                 err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
382                 if (err)
383                         goto failed;
384         }
385         return 0;
386 failed:
387         for (; i >= 0; i--) {
388                 /*
389                  * We want to ext4_forget() only freshly allocated indirect
390                  * blocks.  Buffer for new_blocks[i-1] is at branch[i].bh and
391                  * buffer at branch[0].bh is indirect block / inode already
392                  * existing before ext4_alloc_branch() was called.
393                  */
394                 if (i > 0 && i != indirect_blks && branch[i].bh)
395                         ext4_forget(handle, 1, ar->inode, branch[i].bh,
396                                     branch[i].bh->b_blocknr);
397                 ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
398                                  (i == indirect_blks) ? ar->len : 1, 0);
399         }
400         return err;
401 }
402
403 /**
404  * ext4_splice_branch() - splice the allocated branch onto inode.
405  * @handle: handle for this transaction
406  * @ar: structure describing the allocation request
407  * @where: location of missing link
408  * @num:   number of indirect blocks we are adding
409  *
410  * This function fills the missing link and does all housekeeping needed in
411  * inode (->i_blocks, etc.). In case of success we end up with the full
412  * chain to new block and return 0.
413  */
414 static int ext4_splice_branch(handle_t *handle,
415                               struct ext4_allocation_request *ar,
416                               Indirect *where, int num)
417 {
418         int i;
419         int err = 0;
420         ext4_fsblk_t current_block;
421
422         /*
423          * If we're splicing into a [td]indirect block (as opposed to the
424          * inode) then we need to get write access to the [td]indirect block
425          * before the splice.
426          */
427         if (where->bh) {
428                 BUFFER_TRACE(where->bh, "get_write_access");
429                 err = ext4_journal_get_write_access(handle, where->bh);
430                 if (err)
431                         goto err_out;
432         }
433         /* That's it */
434
435         *where->p = where->key;
436
437         /*
438          * Update the host buffer_head or inode to point to more just allocated
439          * direct blocks blocks
440          */
441         if (num == 0 && ar->len > 1) {
442                 current_block = le32_to_cpu(where->key) + 1;
443                 for (i = 1; i < ar->len; i++)
444                         *(where->p + i) = cpu_to_le32(current_block++);
445         }
446
447         /* We are done with atomic stuff, now do the rest of housekeeping */
448         /* had we spliced it onto indirect block? */
449         if (where->bh) {
450                 /*
451                  * If we spliced it onto an indirect block, we haven't
452                  * altered the inode.  Note however that if it is being spliced
453                  * onto an indirect block at the very end of the file (the
454                  * file is growing) then we *will* alter the inode to reflect
455                  * the new i_size.  But that is not done here - it is done in
456                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
457                  */
458                 jbd_debug(5, "splicing indirect only\n");
459                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
460                 err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
461                 if (err)
462                         goto err_out;
463         } else {
464                 /*
465                  * OK, we spliced it into the inode itself on a direct block.
466                  */
467                 ext4_mark_inode_dirty(handle, ar->inode);
468                 jbd_debug(5, "splicing direct\n");
469         }
470         return err;
471
472 err_out:
473         for (i = 1; i <= num; i++) {
474                 /*
475                  * branch[i].bh is newly allocated, so there is no
476                  * need to revoke the block, which is why we don't
477                  * need to set EXT4_FREE_BLOCKS_METADATA.
478                  */
479                 ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
480                                  EXT4_FREE_BLOCKS_FORGET);
481         }
482         ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
483                          ar->len, 0);
484
485         return err;
486 }
487
488 /*
489  * The ext4_ind_map_blocks() function handles non-extents inodes
490  * (i.e., using the traditional indirect/double-indirect i_blocks
491  * scheme) for ext4_map_blocks().
492  *
493  * Allocation strategy is simple: if we have to allocate something, we will
494  * have to go the whole way to leaf. So let's do it before attaching anything
495  * to tree, set linkage between the newborn blocks, write them if sync is
496  * required, recheck the path, free and repeat if check fails, otherwise
497  * set the last missing link (that will protect us from any truncate-generated
498  * removals - all blocks on the path are immune now) and possibly force the
499  * write on the parent block.
500  * That has a nice additional property: no special recovery from the failed
501  * allocations is needed - we simply release blocks and do not touch anything
502  * reachable from inode.
503  *
504  * `handle' can be NULL if create == 0.
505  *
506  * return > 0, # of blocks mapped or allocated.
507  * return = 0, if plain lookup failed.
508  * return < 0, error case.
509  *
510  * The ext4_ind_get_blocks() function should be called with
511  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
512  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
513  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
514  * blocks.
515  */
516 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
517                         struct ext4_map_blocks *map,
518                         int flags)
519 {
520         struct ext4_allocation_request ar;
521         int err = -EIO;
522         ext4_lblk_t offsets[4];
523         Indirect chain[4];
524         Indirect *partial;
525         int indirect_blks;
526         int blocks_to_boundary = 0;
527         int depth;
528         int count = 0;
529         ext4_fsblk_t first_block = 0;
530
531         trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
532         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
533         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
534         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
535                                    &blocks_to_boundary);
536
537         if (depth == 0)
538                 goto out;
539
540         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
541
542         /* Simplest case - block found, no allocation needed */
543         if (!partial) {
544                 first_block = le32_to_cpu(chain[depth - 1].key);
545                 count++;
546                 /*map more blocks*/
547                 while (count < map->m_len && count <= blocks_to_boundary) {
548                         ext4_fsblk_t blk;
549
550                         blk = le32_to_cpu(*(chain[depth-1].p + count));
551
552                         if (blk == first_block + count)
553                                 count++;
554                         else
555                                 break;
556                 }
557                 goto got_it;
558         }
559
560         /* Next simple case - plain lookup failed */
561         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
562                 unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
563                 int i;
564
565                 /*
566                  * Count number blocks in a subtree under 'partial'. At each
567                  * level we count number of complete empty subtrees beyond
568                  * current offset and then descend into the subtree only
569                  * partially beyond current offset.
570                  */
571                 count = 0;
572                 for (i = partial - chain + 1; i < depth; i++)
573                         count = count * epb + (epb - offsets[i] - 1);
574                 count++;
575                 /* Fill in size of a hole we found */
576                 map->m_pblk = 0;
577                 map->m_len = min_t(unsigned int, map->m_len, count);
578                 goto cleanup;
579         }
580
581         /* Failed read of indirect block */
582         if (err == -EIO)
583                 goto cleanup;
584
585         /*
586          * Okay, we need to do block allocation.
587         */
588         if (ext4_has_feature_bigalloc(inode->i_sb)) {
589                 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
590                                  "non-extent mapped inodes with bigalloc");
591                 return -EFSCORRUPTED;
592         }
593
594         /* Set up for the direct block allocation */
595         memset(&ar, 0, sizeof(ar));
596         ar.inode = inode;
597         ar.logical = map->m_lblk;
598         if (S_ISREG(inode->i_mode))
599                 ar.flags = EXT4_MB_HINT_DATA;
600         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
601                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
602         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
603                 ar.flags |= EXT4_MB_USE_RESERVED;
604
605         ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
606
607         /* the number of blocks need to allocate for [d,t]indirect blocks */
608         indirect_blks = (chain + depth) - partial - 1;
609
610         /*
611          * Next look up the indirect map to count the totoal number of
612          * direct blocks to allocate for this branch.
613          */
614         ar.len = ext4_blks_to_allocate(partial, indirect_blks,
615                                        map->m_len, blocks_to_boundary);
616
617         /*
618          * Block out ext4_truncate while we alter the tree
619          */
620         err = ext4_alloc_branch(handle, &ar, indirect_blks,
621                                 offsets + (partial - chain), partial);
622
623         /*
624          * The ext4_splice_branch call will free and forget any buffers
625          * on the new chain if there is a failure, but that risks using
626          * up transaction credits, especially for bitmaps where the
627          * credits cannot be returned.  Can we handle this somehow?  We
628          * may need to return -EAGAIN upwards in the worst case.  --sct
629          */
630         if (!err)
631                 err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
632         if (err)
633                 goto cleanup;
634
635         map->m_flags |= EXT4_MAP_NEW;
636
637         ext4_update_inode_fsync_trans(handle, inode, 1);
638         count = ar.len;
639
640         /*
641          * Update reserved blocks/metadata blocks after successful block
642          * allocation which had been deferred till now.
643          */
644         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
645                 ext4_da_update_reserve_space(inode, count, 1);
646
647 got_it:
648         map->m_flags |= EXT4_MAP_MAPPED;
649         map->m_pblk = le32_to_cpu(chain[depth-1].key);
650         map->m_len = count;
651         if (count > blocks_to_boundary)
652                 map->m_flags |= EXT4_MAP_BOUNDARY;
653         err = count;
654         /* Clean up and exit */
655         partial = chain + depth - 1;    /* the whole chain */
656 cleanup:
657         while (partial > chain) {
658                 BUFFER_TRACE(partial->bh, "call brelse");
659                 brelse(partial->bh);
660                 partial--;
661         }
662 out:
663         trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
664         return err;
665 }
666
667 /*
668  * Calculate the number of metadata blocks need to reserve
669  * to allocate a new block at @lblocks for non extent file based file
670  */
671 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
672 {
673         struct ext4_inode_info *ei = EXT4_I(inode);
674         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
675         int blk_bits;
676
677         if (lblock < EXT4_NDIR_BLOCKS)
678                 return 0;
679
680         lblock -= EXT4_NDIR_BLOCKS;
681
682         if (ei->i_da_metadata_calc_len &&
683             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
684                 ei->i_da_metadata_calc_len++;
685                 return 0;
686         }
687         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
688         ei->i_da_metadata_calc_len = 1;
689         blk_bits = order_base_2(lblock);
690         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
691 }
692
693 /*
694  * Calculate number of indirect blocks touched by mapping @nrblocks logically
695  * contiguous blocks
696  */
697 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
698 {
699         /*
700          * With N contiguous data blocks, we need at most
701          * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
702          * 2 dindirect blocks, and 1 tindirect block
703          */
704         return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
705 }
706
707 /*
708  * Truncate transactions can be complex and absolutely huge.  So we need to
709  * be able to restart the transaction at a conventient checkpoint to make
710  * sure we don't overflow the journal.
711  *
712  * Try to extend this transaction for the purposes of truncation.  If
713  * extend fails, we need to propagate the failure up and restart the
714  * transaction in the top-level truncate loop. --sct
715  *
716  * Returns 0 if we managed to create more room.  If we can't create more
717  * room, and the transaction must be restarted we return 1.
718  */
719 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
720 {
721         if (!ext4_handle_valid(handle))
722                 return 0;
723         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
724                 return 0;
725         if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
726                 return 0;
727         return 1;
728 }
729
730 /*
731  * Probably it should be a library function... search for first non-zero word
732  * or memcmp with zero_page, whatever is better for particular architecture.
733  * Linus?
734  */
735 static inline int all_zeroes(__le32 *p, __le32 *q)
736 {
737         while (p < q)
738                 if (*p++)
739                         return 0;
740         return 1;
741 }
742
743 /**
744  *      ext4_find_shared - find the indirect blocks for partial truncation.
745  *      @inode:   inode in question
746  *      @depth:   depth of the affected branch
747  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
748  *      @chain:   place to store the pointers to partial indirect blocks
749  *      @top:     place to the (detached) top of branch
750  *
751  *      This is a helper function used by ext4_truncate().
752  *
753  *      When we do truncate() we may have to clean the ends of several
754  *      indirect blocks but leave the blocks themselves alive. Block is
755  *      partially truncated if some data below the new i_size is referred
756  *      from it (and it is on the path to the first completely truncated
757  *      data block, indeed).  We have to free the top of that path along
758  *      with everything to the right of the path. Since no allocation
759  *      past the truncation point is possible until ext4_truncate()
760  *      finishes, we may safely do the latter, but top of branch may
761  *      require special attention - pageout below the truncation point
762  *      might try to populate it.
763  *
764  *      We atomically detach the top of branch from the tree, store the
765  *      block number of its root in *@top, pointers to buffer_heads of
766  *      partially truncated blocks - in @chain[].bh and pointers to
767  *      their last elements that should not be removed - in
768  *      @chain[].p. Return value is the pointer to last filled element
769  *      of @chain.
770  *
771  *      The work left to caller to do the actual freeing of subtrees:
772  *              a) free the subtree starting from *@top
773  *              b) free the subtrees whose roots are stored in
774  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
775  *              c) free the subtrees growing from the inode past the @chain[0].
776  *                      (no partially truncated stuff there).  */
777
778 static Indirect *ext4_find_shared(struct inode *inode, int depth,
779                                   ext4_lblk_t offsets[4], Indirect chain[4],
780                                   __le32 *top)
781 {
782         Indirect *partial, *p;
783         int k, err;
784
785         *top = 0;
786         /* Make k index the deepest non-null offset + 1 */
787         for (k = depth; k > 1 && !offsets[k-1]; k--)
788                 ;
789         partial = ext4_get_branch(inode, k, offsets, chain, &err);
790         /* Writer: pointers */
791         if (!partial)
792                 partial = chain + k-1;
793         /*
794          * If the branch acquired continuation since we've looked at it -
795          * fine, it should all survive and (new) top doesn't belong to us.
796          */
797         if (!partial->key && *partial->p)
798                 /* Writer: end */
799                 goto no_top;
800         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
801                 ;
802         /*
803          * OK, we've found the last block that must survive. The rest of our
804          * branch should be detached before unlocking. However, if that rest
805          * of branch is all ours and does not grow immediately from the inode
806          * it's easier to cheat and just decrement partial->p.
807          */
808         if (p == chain + k - 1 && p > chain) {
809                 p->p--;
810         } else {
811                 *top = *p->p;
812                 /* Nope, don't do this in ext4.  Must leave the tree intact */
813 #if 0
814                 *p->p = 0;
815 #endif
816         }
817         /* Writer: end */
818
819         while (partial > p) {
820                 brelse(partial->bh);
821                 partial--;
822         }
823 no_top:
824         return partial;
825 }
826
827 /*
828  * Zero a number of block pointers in either an inode or an indirect block.
829  * If we restart the transaction we must again get write access to the
830  * indirect block for further modification.
831  *
832  * We release `count' blocks on disk, but (last - first) may be greater
833  * than `count' because there can be holes in there.
834  *
835  * Return 0 on success, 1 on invalid block range
836  * and < 0 on fatal error.
837  */
838 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
839                              struct buffer_head *bh,
840                              ext4_fsblk_t block_to_free,
841                              unsigned long count, __le32 *first,
842                              __le32 *last)
843 {
844         __le32 *p;
845         int     flags = EXT4_FREE_BLOCKS_VALIDATED;
846         int     err;
847
848         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
849             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
850                 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
851         else if (ext4_should_journal_data(inode))
852                 flags |= EXT4_FREE_BLOCKS_FORGET;
853
854         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
855                                    count)) {
856                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
857                                  "blocks %llu len %lu",
858                                  (unsigned long long) block_to_free, count);
859                 return 1;
860         }
861
862         if (try_to_extend_transaction(handle, inode)) {
863                 if (bh) {
864                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
865                         err = ext4_handle_dirty_metadata(handle, inode, bh);
866                         if (unlikely(err))
867                                 goto out_err;
868                 }
869                 err = ext4_mark_inode_dirty(handle, inode);
870                 if (unlikely(err))
871                         goto out_err;
872                 err = ext4_truncate_restart_trans(handle, inode,
873                                         ext4_blocks_for_truncate(inode));
874                 if (unlikely(err))
875                         goto out_err;
876                 if (bh) {
877                         BUFFER_TRACE(bh, "retaking write access");
878                         err = ext4_journal_get_write_access(handle, bh);
879                         if (unlikely(err))
880                                 goto out_err;
881                 }
882         }
883
884         for (p = first; p < last; p++)
885                 *p = 0;
886
887         ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
888         return 0;
889 out_err:
890         ext4_std_error(inode->i_sb, err);
891         return err;
892 }
893
894 /**
895  * ext4_free_data - free a list of data blocks
896  * @handle:     handle for this transaction
897  * @inode:      inode we are dealing with
898  * @this_bh:    indirect buffer_head which contains *@first and *@last
899  * @first:      array of block numbers
900  * @last:       points immediately past the end of array
901  *
902  * We are freeing all blocks referred from that array (numbers are stored as
903  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
904  *
905  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
906  * blocks are contiguous then releasing them at one time will only affect one
907  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
908  * actually use a lot of journal space.
909  *
910  * @this_bh will be %NULL if @first and @last point into the inode's direct
911  * block pointers.
912  */
913 static void ext4_free_data(handle_t *handle, struct inode *inode,
914                            struct buffer_head *this_bh,
915                            __le32 *first, __le32 *last)
916 {
917         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
918         unsigned long count = 0;            /* Number of blocks in the run */
919         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
920                                                corresponding to
921                                                block_to_free */
922         ext4_fsblk_t nr;                    /* Current block # */
923         __le32 *p;                          /* Pointer into inode/ind
924                                                for current block */
925         int err = 0;
926
927         if (this_bh) {                          /* For indirect block */
928                 BUFFER_TRACE(this_bh, "get_write_access");
929                 err = ext4_journal_get_write_access(handle, this_bh);
930                 /* Important: if we can't update the indirect pointers
931                  * to the blocks, we can't free them. */
932                 if (err)
933                         return;
934         }
935
936         for (p = first; p < last; p++) {
937                 nr = le32_to_cpu(*p);
938                 if (nr) {
939                         /* accumulate blocks to free if they're contiguous */
940                         if (count == 0) {
941                                 block_to_free = nr;
942                                 block_to_free_p = p;
943                                 count = 1;
944                         } else if (nr == block_to_free + count) {
945                                 count++;
946                         } else {
947                                 err = ext4_clear_blocks(handle, inode, this_bh,
948                                                         block_to_free, count,
949                                                         block_to_free_p, p);
950                                 if (err)
951                                         break;
952                                 block_to_free = nr;
953                                 block_to_free_p = p;
954                                 count = 1;
955                         }
956                 }
957         }
958
959         if (!err && count > 0)
960                 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
961                                         count, block_to_free_p, p);
962         if (err < 0)
963                 /* fatal error */
964                 return;
965
966         if (this_bh) {
967                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
968
969                 /*
970                  * The buffer head should have an attached journal head at this
971                  * point. However, if the data is corrupted and an indirect
972                  * block pointed to itself, it would have been detached when
973                  * the block was cleared. Check for this instead of OOPSing.
974                  */
975                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
976                         ext4_handle_dirty_metadata(handle, inode, this_bh);
977                 else
978                         EXT4_ERROR_INODE(inode,
979                                          "circular indirect block detected at "
980                                          "block %llu",
981                                 (unsigned long long) this_bh->b_blocknr);
982         }
983 }
984
985 /**
986  *      ext4_free_branches - free an array of branches
987  *      @handle: JBD handle for this transaction
988  *      @inode: inode we are dealing with
989  *      @parent_bh: the buffer_head which contains *@first and *@last
990  *      @first: array of block numbers
991  *      @last:  pointer immediately past the end of array
992  *      @depth: depth of the branches to free
993  *
994  *      We are freeing all blocks referred from these branches (numbers are
995  *      stored as little-endian 32-bit) and updating @inode->i_blocks
996  *      appropriately.
997  */
998 static void ext4_free_branches(handle_t *handle, struct inode *inode,
999                                struct buffer_head *parent_bh,
1000                                __le32 *first, __le32 *last, int depth)
1001 {
1002         ext4_fsblk_t nr;
1003         __le32 *p;
1004
1005         if (ext4_handle_is_aborted(handle))
1006                 return;
1007
1008         if (depth--) {
1009                 struct buffer_head *bh;
1010                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1011                 p = last;
1012                 while (--p >= first) {
1013                         nr = le32_to_cpu(*p);
1014                         if (!nr)
1015                                 continue;               /* A hole */
1016
1017                         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1018                                                    nr, 1)) {
1019                                 EXT4_ERROR_INODE(inode,
1020                                                  "invalid indirect mapped "
1021                                                  "block %lu (level %d)",
1022                                                  (unsigned long) nr, depth);
1023                                 break;
1024                         }
1025
1026                         /* Go read the buffer for the next level down */
1027                         bh = sb_bread(inode->i_sb, nr);
1028
1029                         /*
1030                          * A read failure? Report error and clear slot
1031                          * (should be rare).
1032                          */
1033                         if (!bh) {
1034                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
1035                                                        "Read failure");
1036                                 continue;
1037                         }
1038
1039                         /* This zaps the entire block.  Bottom up. */
1040                         BUFFER_TRACE(bh, "free child branches");
1041                         ext4_free_branches(handle, inode, bh,
1042                                         (__le32 *) bh->b_data,
1043                                         (__le32 *) bh->b_data + addr_per_block,
1044                                         depth);
1045                         brelse(bh);
1046
1047                         /*
1048                          * Everything below this this pointer has been
1049                          * released.  Now let this top-of-subtree go.
1050                          *
1051                          * We want the freeing of this indirect block to be
1052                          * atomic in the journal with the updating of the
1053                          * bitmap block which owns it.  So make some room in
1054                          * the journal.
1055                          *
1056                          * We zero the parent pointer *after* freeing its
1057                          * pointee in the bitmaps, so if extend_transaction()
1058                          * for some reason fails to put the bitmap changes and
1059                          * the release into the same transaction, recovery
1060                          * will merely complain about releasing a free block,
1061                          * rather than leaking blocks.
1062                          */
1063                         if (ext4_handle_is_aborted(handle))
1064                                 return;
1065                         if (try_to_extend_transaction(handle, inode)) {
1066                                 ext4_mark_inode_dirty(handle, inode);
1067                                 ext4_truncate_restart_trans(handle, inode,
1068                                             ext4_blocks_for_truncate(inode));
1069                         }
1070
1071                         /*
1072                          * The forget flag here is critical because if
1073                          * we are journaling (and not doing data
1074                          * journaling), we have to make sure a revoke
1075                          * record is written to prevent the journal
1076                          * replay from overwriting the (former)
1077                          * indirect block if it gets reallocated as a
1078                          * data block.  This must happen in the same
1079                          * transaction where the data blocks are
1080                          * actually freed.
1081                          */
1082                         ext4_free_blocks(handle, inode, NULL, nr, 1,
1083                                          EXT4_FREE_BLOCKS_METADATA|
1084                                          EXT4_FREE_BLOCKS_FORGET);
1085
1086                         if (parent_bh) {
1087                                 /*
1088                                  * The block which we have just freed is
1089                                  * pointed to by an indirect block: journal it
1090                                  */
1091                                 BUFFER_TRACE(parent_bh, "get_write_access");
1092                                 if (!ext4_journal_get_write_access(handle,
1093                                                                    parent_bh)){
1094                                         *p = 0;
1095                                         BUFFER_TRACE(parent_bh,
1096                                         "call ext4_handle_dirty_metadata");
1097                                         ext4_handle_dirty_metadata(handle,
1098                                                                    inode,
1099                                                                    parent_bh);
1100                                 }
1101                         }
1102                 }
1103         } else {
1104                 /* We have reached the bottom of the tree. */
1105                 BUFFER_TRACE(parent_bh, "free data blocks");
1106                 ext4_free_data(handle, inode, parent_bh, first, last);
1107         }
1108 }
1109
1110 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1111 {
1112         struct ext4_inode_info *ei = EXT4_I(inode);
1113         __le32 *i_data = ei->i_data;
1114         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1115         ext4_lblk_t offsets[4];
1116         Indirect chain[4];
1117         Indirect *partial;
1118         __le32 nr = 0;
1119         int n = 0;
1120         ext4_lblk_t last_block, max_block;
1121         unsigned blocksize = inode->i_sb->s_blocksize;
1122
1123         last_block = (inode->i_size + blocksize-1)
1124                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1125         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1126                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1127
1128         if (last_block != max_block) {
1129                 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1130                 if (n == 0)
1131                         return;
1132         }
1133
1134         ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1135
1136         /*
1137          * The orphan list entry will now protect us from any crash which
1138          * occurs before the truncate completes, so it is now safe to propagate
1139          * the new, shorter inode size (held for now in i_size) into the
1140          * on-disk inode. We do this via i_disksize, which is the value which
1141          * ext4 *really* writes onto the disk inode.
1142          */
1143         ei->i_disksize = inode->i_size;
1144
1145         if (last_block == max_block) {
1146                 /*
1147                  * It is unnecessary to free any data blocks if last_block is
1148                  * equal to the indirect block limit.
1149                  */
1150                 return;
1151         } else if (n == 1) {            /* direct blocks */
1152                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1153                                i_data + EXT4_NDIR_BLOCKS);
1154                 goto do_indirects;
1155         }
1156
1157         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1158         /* Kill the top of shared branch (not detached) */
1159         if (nr) {
1160                 if (partial == chain) {
1161                         /* Shared branch grows from the inode */
1162                         ext4_free_branches(handle, inode, NULL,
1163                                            &nr, &nr+1, (chain+n-1) - partial);
1164                         *partial->p = 0;
1165                         /*
1166                          * We mark the inode dirty prior to restart,
1167                          * and prior to stop.  No need for it here.
1168                          */
1169                 } else {
1170                         /* Shared branch grows from an indirect block */
1171                         BUFFER_TRACE(partial->bh, "get_write_access");
1172                         ext4_free_branches(handle, inode, partial->bh,
1173                                         partial->p,
1174                                         partial->p+1, (chain+n-1) - partial);
1175                 }
1176         }
1177         /* Clear the ends of indirect blocks on the shared branch */
1178         while (partial > chain) {
1179                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1180                                    (__le32*)partial->bh->b_data+addr_per_block,
1181                                    (chain+n-1) - partial);
1182                 BUFFER_TRACE(partial->bh, "call brelse");
1183                 brelse(partial->bh);
1184                 partial--;
1185         }
1186 do_indirects:
1187         /* Kill the remaining (whole) subtrees */
1188         switch (offsets[0]) {
1189         default:
1190                 nr = i_data[EXT4_IND_BLOCK];
1191                 if (nr) {
1192                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1193                         i_data[EXT4_IND_BLOCK] = 0;
1194                 }
1195                 /* fall through */
1196         case EXT4_IND_BLOCK:
1197                 nr = i_data[EXT4_DIND_BLOCK];
1198                 if (nr) {
1199                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1200                         i_data[EXT4_DIND_BLOCK] = 0;
1201                 }
1202                 /* fall through */
1203         case EXT4_DIND_BLOCK:
1204                 nr = i_data[EXT4_TIND_BLOCK];
1205                 if (nr) {
1206                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1207                         i_data[EXT4_TIND_BLOCK] = 0;
1208                 }
1209                 /* fall through */
1210         case EXT4_TIND_BLOCK:
1211                 ;
1212         }
1213 }
1214
1215 /**
1216  *      ext4_ind_remove_space - remove space from the range
1217  *      @handle: JBD handle for this transaction
1218  *      @inode: inode we are dealing with
1219  *      @start: First block to remove
1220  *      @end:   One block after the last block to remove (exclusive)
1221  *
1222  *      Free the blocks in the defined range (end is exclusive endpoint of
1223  *      range). This is used by ext4_punch_hole().
1224  */
1225 int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1226                           ext4_lblk_t start, ext4_lblk_t end)
1227 {
1228         struct ext4_inode_info *ei = EXT4_I(inode);
1229         __le32 *i_data = ei->i_data;
1230         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1231         ext4_lblk_t offsets[4], offsets2[4];
1232         Indirect chain[4], chain2[4];
1233         Indirect *partial, *partial2;
1234         Indirect *p = NULL, *p2 = NULL;
1235         ext4_lblk_t max_block;
1236         __le32 nr = 0, nr2 = 0;
1237         int n = 0, n2 = 0;
1238         unsigned blocksize = inode->i_sb->s_blocksize;
1239
1240         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1241                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1242         if (end >= max_block)
1243                 end = max_block;
1244         if ((start >= end) || (start > max_block))
1245                 return 0;
1246
1247         n = ext4_block_to_path(inode, start, offsets, NULL);
1248         n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1249
1250         BUG_ON(n > n2);
1251
1252         if ((n == 1) && (n == n2)) {
1253                 /* We're punching only within direct block range */
1254                 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1255                                i_data + offsets2[0]);
1256                 return 0;
1257         } else if (n2 > n) {
1258                 /*
1259                  * Start and end are on a different levels so we're going to
1260                  * free partial block at start, and partial block at end of
1261                  * the range. If there are some levels in between then
1262                  * do_indirects label will take care of that.
1263                  */
1264
1265                 if (n == 1) {
1266                         /*
1267                          * Start is at the direct block level, free
1268                          * everything to the end of the level.
1269                          */
1270                         ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1271                                        i_data + EXT4_NDIR_BLOCKS);
1272                         goto end_range;
1273                 }
1274
1275
1276                 partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1277                 if (nr) {
1278                         if (partial == chain) {
1279                                 /* Shared branch grows from the inode */
1280                                 ext4_free_branches(handle, inode, NULL,
1281                                            &nr, &nr+1, (chain+n-1) - partial);
1282                                 *partial->p = 0;
1283                         } else {
1284                                 /* Shared branch grows from an indirect block */
1285                                 BUFFER_TRACE(partial->bh, "get_write_access");
1286                                 ext4_free_branches(handle, inode, partial->bh,
1287                                         partial->p,
1288                                         partial->p+1, (chain+n-1) - partial);
1289                         }
1290                 }
1291
1292                 /*
1293                  * Clear the ends of indirect blocks on the shared branch
1294                  * at the start of the range
1295                  */
1296                 while (partial > chain) {
1297                         ext4_free_branches(handle, inode, partial->bh,
1298                                 partial->p + 1,
1299                                 (__le32 *)partial->bh->b_data+addr_per_block,
1300                                 (chain+n-1) - partial);
1301                         partial--;
1302                 }
1303
1304 end_range:
1305                 partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1306                 if (nr2) {
1307                         if (partial2 == chain2) {
1308                                 /*
1309                                  * Remember, end is exclusive so here we're at
1310                                  * the start of the next level we're not going
1311                                  * to free. Everything was covered by the start
1312                                  * of the range.
1313                                  */
1314                                 goto do_indirects;
1315                         }
1316                 } else {
1317                         /*
1318                          * ext4_find_shared returns Indirect structure which
1319                          * points to the last element which should not be
1320                          * removed by truncate. But this is end of the range
1321                          * in punch_hole so we need to point to the next element
1322                          */
1323                         partial2->p++;
1324                 }
1325
1326                 /*
1327                  * Clear the ends of indirect blocks on the shared branch
1328                  * at the end of the range
1329                  */
1330                 while (partial2 > chain2) {
1331                         ext4_free_branches(handle, inode, partial2->bh,
1332                                            (__le32 *)partial2->bh->b_data,
1333                                            partial2->p,
1334                                            (chain2+n2-1) - partial2);
1335                         partial2--;
1336                 }
1337                 goto do_indirects;
1338         }
1339
1340         /* Punch happened within the same level (n == n2) */
1341         partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1342         partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1343
1344         /* Free top, but only if partial2 isn't its subtree. */
1345         if (nr) {
1346                 int level = min(partial - chain, partial2 - chain2);
1347                 int i;
1348                 int subtree = 1;
1349
1350                 for (i = 0; i <= level; i++) {
1351                         if (offsets[i] != offsets2[i]) {
1352                                 subtree = 0;
1353                                 break;
1354                         }
1355                 }
1356
1357                 if (!subtree) {
1358                         if (partial == chain) {
1359                                 /* Shared branch grows from the inode */
1360                                 ext4_free_branches(handle, inode, NULL,
1361                                                    &nr, &nr+1,
1362                                                    (chain+n-1) - partial);
1363                                 *partial->p = 0;
1364                         } else {
1365                                 /* Shared branch grows from an indirect block */
1366                                 BUFFER_TRACE(partial->bh, "get_write_access");
1367                                 ext4_free_branches(handle, inode, partial->bh,
1368                                                    partial->p,
1369                                                    partial->p+1,
1370                                                    (chain+n-1) - partial);
1371                         }
1372                 }
1373         }
1374
1375         if (!nr2) {
1376                 /*
1377                  * ext4_find_shared returns Indirect structure which
1378                  * points to the last element which should not be
1379                  * removed by truncate. But this is end of the range
1380                  * in punch_hole so we need to point to the next element
1381                  */
1382                 partial2->p++;
1383         }
1384
1385         while (partial > chain || partial2 > chain2) {
1386                 int depth = (chain+n-1) - partial;
1387                 int depth2 = (chain2+n2-1) - partial2;
1388
1389                 if (partial > chain && partial2 > chain2 &&
1390                     partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1391                         /*
1392                          * We've converged on the same block. Clear the range,
1393                          * then we're done.
1394                          */
1395                         ext4_free_branches(handle, inode, partial->bh,
1396                                            partial->p + 1,
1397                                            partial2->p,
1398                                            (chain+n-1) - partial);
1399                         goto cleanup;
1400                 }
1401
1402                 /*
1403                  * The start and end partial branches may not be at the same
1404                  * level even though the punch happened within one level. So, we
1405                  * give them a chance to arrive at the same level, then walk
1406                  * them in step with each other until we converge on the same
1407                  * block.
1408                  */
1409                 if (partial > chain && depth <= depth2) {
1410                         ext4_free_branches(handle, inode, partial->bh,
1411                                            partial->p + 1,
1412                                            (__le32 *)partial->bh->b_data+addr_per_block,
1413                                            (chain+n-1) - partial);
1414                         partial--;
1415                 }
1416                 if (partial2 > chain2 && depth2 <= depth) {
1417                         ext4_free_branches(handle, inode, partial2->bh,
1418                                            (__le32 *)partial2->bh->b_data,
1419                                            partial2->p,
1420                                            (chain2+n2-1) - partial2);
1421                         partial2--;
1422                 }
1423         }
1424
1425 cleanup:
1426         while (p && p > chain) {
1427                 BUFFER_TRACE(p->bh, "call brelse");
1428                 brelse(p->bh);
1429                 p--;
1430         }
1431         while (p2 && p2 > chain2) {
1432                 BUFFER_TRACE(p2->bh, "call brelse");
1433                 brelse(p2->bh);
1434                 p2--;
1435         }
1436         return 0;
1437
1438 do_indirects:
1439         /* Kill the remaining (whole) subtrees */
1440         switch (offsets[0]) {
1441         default:
1442                 if (++n >= n2)
1443                         break;
1444                 nr = i_data[EXT4_IND_BLOCK];
1445                 if (nr) {
1446                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1447                         i_data[EXT4_IND_BLOCK] = 0;
1448                 }
1449                 /* fall through */
1450         case EXT4_IND_BLOCK:
1451                 if (++n >= n2)
1452                         break;
1453                 nr = i_data[EXT4_DIND_BLOCK];
1454                 if (nr) {
1455                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1456                         i_data[EXT4_DIND_BLOCK] = 0;
1457                 }
1458                 /* fall through */
1459         case EXT4_DIND_BLOCK:
1460                 if (++n >= n2)
1461                         break;
1462                 nr = i_data[EXT4_TIND_BLOCK];
1463                 if (nr) {
1464                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1465                         i_data[EXT4_TIND_BLOCK] = 0;
1466                 }
1467                 /* fall through */
1468         case EXT4_TIND_BLOCK:
1469                 ;
1470         }
1471         goto cleanup;
1472 }