GNU Linux-libre 4.19.211-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         int ret = -EIO;
152
153         *err = 0;
154         /* i_data is not going away, no lock needed */
155         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
156         if (!p->key)
157                 goto no_block;
158         while (--depth) {
159                 bh = sb_getblk(sb, le32_to_cpu(p->key));
160                 if (unlikely(!bh)) {
161                         ret = -ENOMEM;
162                         goto failure;
163                 }
164
165                 if (!bh_uptodate_or_lock(bh)) {
166                         if (bh_submit_read(bh) < 0) {
167                                 put_bh(bh);
168                                 goto failure;
169                         }
170                         /* validate block references */
171                         if (ext4_check_indirect_blockref(inode, bh)) {
172                                 put_bh(bh);
173                                 goto failure;
174                         }
175                 }
176
177                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
178                 /* Reader: end */
179                 if (!p->key)
180                         goto no_block;
181         }
182         return NULL;
183
184 failure:
185         *err = ret;
186 no_block:
187         return p;
188 }
189
190 /**
191  *      ext4_find_near - find a place for allocation with sufficient locality
192  *      @inode: owner
193  *      @ind: descriptor of indirect block.
194  *
195  *      This function returns the preferred place for block allocation.
196  *      It is used when heuristic for sequential allocation fails.
197  *      Rules are:
198  *        + if there is a block to the left of our position - allocate near it.
199  *        + if pointer will live in indirect block - allocate near that block.
200  *        + if pointer will live in inode - allocate in the same
201  *          cylinder group.
202  *
203  * In the latter case we colour the starting block by the callers PID to
204  * prevent it from clashing with concurrent allocations for a different inode
205  * in the same block group.   The PID is used here so that functionally related
206  * files will be close-by on-disk.
207  *
208  *      Caller must make sure that @ind is valid and will stay that way.
209  */
210 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
211 {
212         struct ext4_inode_info *ei = EXT4_I(inode);
213         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
214         __le32 *p;
215
216         /* Try to find previous block */
217         for (p = ind->p - 1; p >= start; p--) {
218                 if (*p)
219                         return le32_to_cpu(*p);
220         }
221
222         /* No such thing, so let's try location of indirect block */
223         if (ind->bh)
224                 return ind->bh->b_blocknr;
225
226         /*
227          * It is going to be referred to from the inode itself? OK, just put it
228          * into the same cylinder group then.
229          */
230         return ext4_inode_to_goal_block(inode);
231 }
232
233 /**
234  *      ext4_find_goal - find a preferred place for allocation.
235  *      @inode: owner
236  *      @block:  block we want
237  *      @partial: pointer to the last triple within a chain
238  *
239  *      Normally this function find the preferred place for block allocation,
240  *      returns it.
241  *      Because this is only used for non-extent files, we limit the block nr
242  *      to 32 bits.
243  */
244 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
245                                    Indirect *partial)
246 {
247         ext4_fsblk_t goal;
248
249         /*
250          * XXX need to get goal block from mballoc's data structures
251          */
252
253         goal = ext4_find_near(inode, partial);
254         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
255         return goal;
256 }
257
258 /**
259  *      ext4_blks_to_allocate - Look up the block map and count the number
260  *      of direct blocks need to be allocated for the given branch.
261  *
262  *      @branch: chain of indirect blocks
263  *      @k: number of blocks need for indirect blocks
264  *      @blks: number of data blocks to be mapped.
265  *      @blocks_to_boundary:  the offset in the indirect block
266  *
267  *      return the total number of blocks to be allocate, including the
268  *      direct and indirect blocks.
269  */
270 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
271                                  int blocks_to_boundary)
272 {
273         unsigned int count = 0;
274
275         /*
276          * Simple case, [t,d]Indirect block(s) has not allocated yet
277          * then it's clear blocks on that path have not allocated
278          */
279         if (k > 0) {
280                 /* right now we don't handle cross boundary allocation */
281                 if (blks < blocks_to_boundary + 1)
282                         count += blks;
283                 else
284                         count += blocks_to_boundary + 1;
285                 return count;
286         }
287
288         count++;
289         while (count < blks && count <= blocks_to_boundary &&
290                 le32_to_cpu(*(branch[0].p + count)) == 0) {
291                 count++;
292         }
293         return count;
294 }
295
296 /**
297  *      ext4_alloc_branch - allocate and set up a chain of blocks.
298  *      @handle: handle for this transaction
299  *      @inode: owner
300  *      @indirect_blks: number of allocated indirect blocks
301  *      @blks: number of allocated direct blocks
302  *      @goal: preferred place for allocation
303  *      @offsets: offsets (in the blocks) to store the pointers to next.
304  *      @branch: place to store the chain in.
305  *
306  *      This function allocates blocks, zeroes out all but the last one,
307  *      links them into chain and (if we are synchronous) writes them to disk.
308  *      In other words, it prepares a branch that can be spliced onto the
309  *      inode. It stores the information about that chain in the branch[], in
310  *      the same format as ext4_get_branch() would do. We are calling it after
311  *      we had read the existing part of chain and partial points to the last
312  *      triple of that (one with zero ->key). Upon the exit we have the same
313  *      picture as after the successful ext4_get_block(), except that in one
314  *      place chain is disconnected - *branch->p is still zero (we did not
315  *      set the last link), but branch->key contains the number that should
316  *      be placed into *branch->p to fill that gap.
317  *
318  *      If allocation fails we free all blocks we've allocated (and forget
319  *      their buffer_heads) and return the error value the from failed
320  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
321  *      as described above and return 0.
322  */
323 static int ext4_alloc_branch(handle_t *handle,
324                              struct ext4_allocation_request *ar,
325                              int indirect_blks, ext4_lblk_t *offsets,
326                              Indirect *branch)
327 {
328         struct buffer_head *            bh;
329         ext4_fsblk_t                    b, new_blocks[4];
330         __le32                          *p;
331         int                             i, j, err, len = 1;
332
333         for (i = 0; i <= indirect_blks; i++) {
334                 if (i == indirect_blks) {
335                         new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
336                 } else
337                         ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
338                                         ar->inode, ar->goal,
339                                         ar->flags & EXT4_MB_DELALLOC_RESERVED,
340                                         NULL, &err);
341                 if (err) {
342                         i--;
343                         goto failed;
344                 }
345                 branch[i].key = cpu_to_le32(new_blocks[i]);
346                 if (i == 0)
347                         continue;
348
349                 bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
350                 if (unlikely(!bh)) {
351                         err = -ENOMEM;
352                         goto failed;
353                 }
354                 lock_buffer(bh);
355                 BUFFER_TRACE(bh, "call get_create_access");
356                 err = ext4_journal_get_create_access(handle, bh);
357                 if (err) {
358                         unlock_buffer(bh);
359                         goto failed;
360                 }
361
362                 memset(bh->b_data, 0, bh->b_size);
363                 p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
364                 b = new_blocks[i];
365
366                 if (i == indirect_blks)
367                         len = ar->len;
368                 for (j = 0; j < len; j++)
369                         *p++ = cpu_to_le32(b++);
370
371                 BUFFER_TRACE(bh, "marking uptodate");
372                 set_buffer_uptodate(bh);
373                 unlock_buffer(bh);
374
375                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
376                 err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
377                 if (err)
378                         goto failed;
379         }
380         return 0;
381 failed:
382         for (; i >= 0; i--) {
383                 /*
384                  * We want to ext4_forget() only freshly allocated indirect
385                  * blocks.  Buffer for new_blocks[i-1] is at branch[i].bh and
386                  * buffer at branch[0].bh is indirect block / inode already
387                  * existing before ext4_alloc_branch() was called.
388                  */
389                 if (i > 0 && i != indirect_blks && branch[i].bh)
390                         ext4_forget(handle, 1, ar->inode, branch[i].bh,
391                                     branch[i].bh->b_blocknr);
392                 ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
393                                  (i == indirect_blks) ? ar->len : 1, 0);
394         }
395         return err;
396 }
397
398 /**
399  * ext4_splice_branch - splice the allocated branch onto inode.
400  * @handle: handle for this transaction
401  * @inode: owner
402  * @block: (logical) number of block we are adding
403  * @chain: chain of indirect blocks (with a missing link - see
404  *      ext4_alloc_branch)
405  * @where: location of missing link
406  * @num:   number of indirect blocks we are adding
407  * @blks:  number of direct blocks we are adding
408  *
409  * This function fills the missing link and does all housekeeping needed in
410  * inode (->i_blocks, etc.). In case of success we end up with the full
411  * chain to new block and return 0.
412  */
413 static int ext4_splice_branch(handle_t *handle,
414                               struct ext4_allocation_request *ar,
415                               Indirect *where, int num)
416 {
417         int i;
418         int err = 0;
419         ext4_fsblk_t current_block;
420
421         /*
422          * If we're splicing into a [td]indirect block (as opposed to the
423          * inode) then we need to get write access to the [td]indirect block
424          * before the splice.
425          */
426         if (where->bh) {
427                 BUFFER_TRACE(where->bh, "get_write_access");
428                 err = ext4_journal_get_write_access(handle, where->bh);
429                 if (err)
430                         goto err_out;
431         }
432         /* That's it */
433
434         *where->p = where->key;
435
436         /*
437          * Update the host buffer_head or inode to point to more just allocated
438          * direct blocks blocks
439          */
440         if (num == 0 && ar->len > 1) {
441                 current_block = le32_to_cpu(where->key) + 1;
442                 for (i = 1; i < ar->len; i++)
443                         *(where->p + i) = cpu_to_le32(current_block++);
444         }
445
446         /* We are done with atomic stuff, now do the rest of housekeeping */
447         /* had we spliced it onto indirect block? */
448         if (where->bh) {
449                 /*
450                  * If we spliced it onto an indirect block, we haven't
451                  * altered the inode.  Note however that if it is being spliced
452                  * onto an indirect block at the very end of the file (the
453                  * file is growing) then we *will* alter the inode to reflect
454                  * the new i_size.  But that is not done here - it is done in
455                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
456                  */
457                 jbd_debug(5, "splicing indirect only\n");
458                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
459                 err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
460                 if (err)
461                         goto err_out;
462         } else {
463                 /*
464                  * OK, we spliced it into the inode itself on a direct block.
465                  */
466                 ext4_mark_inode_dirty(handle, ar->inode);
467                 jbd_debug(5, "splicing direct\n");
468         }
469         return err;
470
471 err_out:
472         for (i = 1; i <= num; i++) {
473                 /*
474                  * branch[i].bh is newly allocated, so there is no
475                  * need to revoke the block, which is why we don't
476                  * need to set EXT4_FREE_BLOCKS_METADATA.
477                  */
478                 ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
479                                  EXT4_FREE_BLOCKS_FORGET);
480         }
481         ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
482                          ar->len, 0);
483
484         return err;
485 }
486
487 /*
488  * The ext4_ind_map_blocks() function handles non-extents inodes
489  * (i.e., using the traditional indirect/double-indirect i_blocks
490  * scheme) for ext4_map_blocks().
491  *
492  * Allocation strategy is simple: if we have to allocate something, we will
493  * have to go the whole way to leaf. So let's do it before attaching anything
494  * to tree, set linkage between the newborn blocks, write them if sync is
495  * required, recheck the path, free and repeat if check fails, otherwise
496  * set the last missing link (that will protect us from any truncate-generated
497  * removals - all blocks on the path are immune now) and possibly force the
498  * write on the parent block.
499  * That has a nice additional property: no special recovery from the failed
500  * allocations is needed - we simply release blocks and do not touch anything
501  * reachable from inode.
502  *
503  * `handle' can be NULL if create == 0.
504  *
505  * return > 0, # of blocks mapped or allocated.
506  * return = 0, if plain lookup failed.
507  * return < 0, error case.
508  *
509  * The ext4_ind_get_blocks() function should be called with
510  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
511  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
512  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
513  * blocks.
514  */
515 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
516                         struct ext4_map_blocks *map,
517                         int flags)
518 {
519         struct ext4_allocation_request ar;
520         int err = -EIO;
521         ext4_lblk_t offsets[4];
522         Indirect chain[4];
523         Indirect *partial;
524         int indirect_blks;
525         int blocks_to_boundary = 0;
526         int depth;
527         int count = 0;
528         ext4_fsblk_t first_block = 0;
529
530         trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
531         J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
532         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
533         depth = ext4_block_to_path(inode, map->m_lblk, offsets,
534                                    &blocks_to_boundary);
535
536         if (depth == 0)
537                 goto out;
538
539         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
540
541         /* Simplest case - block found, no allocation needed */
542         if (!partial) {
543                 first_block = le32_to_cpu(chain[depth - 1].key);
544                 count++;
545                 /*map more blocks*/
546                 while (count < map->m_len && count <= blocks_to_boundary) {
547                         ext4_fsblk_t blk;
548
549                         blk = le32_to_cpu(*(chain[depth-1].p + count));
550
551                         if (blk == first_block + count)
552                                 count++;
553                         else
554                                 break;
555                 }
556                 goto got_it;
557         }
558
559         /* Next simple case - plain lookup failed */
560         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
561                 unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
562                 int i;
563
564                 /*
565                  * Count number blocks in a subtree under 'partial'. At each
566                  * level we count number of complete empty subtrees beyond
567                  * current offset and then descend into the subtree only
568                  * partially beyond current offset.
569                  */
570                 count = 0;
571                 for (i = partial - chain + 1; i < depth; i++)
572                         count = count * epb + (epb - offsets[i] - 1);
573                 count++;
574                 /* Fill in size of a hole we found */
575                 map->m_pblk = 0;
576                 map->m_len = min_t(unsigned int, map->m_len, count);
577                 goto cleanup;
578         }
579
580         /* Failed read of indirect block */
581         if (err == -EIO)
582                 goto cleanup;
583
584         /*
585          * Okay, we need to do block allocation.
586         */
587         if (ext4_has_feature_bigalloc(inode->i_sb)) {
588                 EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
589                                  "non-extent mapped inodes with bigalloc");
590                 return -EFSCORRUPTED;
591         }
592
593         /* Set up for the direct block allocation */
594         memset(&ar, 0, sizeof(ar));
595         ar.inode = inode;
596         ar.logical = map->m_lblk;
597         if (S_ISREG(inode->i_mode))
598                 ar.flags = EXT4_MB_HINT_DATA;
599         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
600                 ar.flags |= EXT4_MB_DELALLOC_RESERVED;
601         if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
602                 ar.flags |= EXT4_MB_USE_RESERVED;
603
604         ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
605
606         /* the number of blocks need to allocate for [d,t]indirect blocks */
607         indirect_blks = (chain + depth) - partial - 1;
608
609         /*
610          * Next look up the indirect map to count the totoal number of
611          * direct blocks to allocate for this branch.
612          */
613         ar.len = ext4_blks_to_allocate(partial, indirect_blks,
614                                        map->m_len, blocks_to_boundary);
615
616         /*
617          * Block out ext4_truncate while we alter the tree
618          */
619         err = ext4_alloc_branch(handle, &ar, indirect_blks,
620                                 offsets + (partial - chain), partial);
621
622         /*
623          * The ext4_splice_branch call will free and forget any buffers
624          * on the new chain if there is a failure, but that risks using
625          * up transaction credits, especially for bitmaps where the
626          * credits cannot be returned.  Can we handle this somehow?  We
627          * may need to return -EAGAIN upwards in the worst case.  --sct
628          */
629         if (!err)
630                 err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
631         if (err)
632                 goto cleanup;
633
634         map->m_flags |= EXT4_MAP_NEW;
635
636         ext4_update_inode_fsync_trans(handle, inode, 1);
637         count = ar.len;
638 got_it:
639         map->m_flags |= EXT4_MAP_MAPPED;
640         map->m_pblk = le32_to_cpu(chain[depth-1].key);
641         map->m_len = count;
642         if (count > blocks_to_boundary)
643                 map->m_flags |= EXT4_MAP_BOUNDARY;
644         err = count;
645         /* Clean up and exit */
646         partial = chain + depth - 1;    /* the whole chain */
647 cleanup:
648         while (partial > chain) {
649                 BUFFER_TRACE(partial->bh, "call brelse");
650                 brelse(partial->bh);
651                 partial--;
652         }
653 out:
654         trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
655         return err;
656 }
657
658 /*
659  * Calculate the number of metadata blocks need to reserve
660  * to allocate a new block at @lblocks for non extent file based file
661  */
662 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
663 {
664         struct ext4_inode_info *ei = EXT4_I(inode);
665         sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
666         int blk_bits;
667
668         if (lblock < EXT4_NDIR_BLOCKS)
669                 return 0;
670
671         lblock -= EXT4_NDIR_BLOCKS;
672
673         if (ei->i_da_metadata_calc_len &&
674             (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
675                 ei->i_da_metadata_calc_len++;
676                 return 0;
677         }
678         ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
679         ei->i_da_metadata_calc_len = 1;
680         blk_bits = order_base_2(lblock);
681         return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
682 }
683
684 /*
685  * Calculate number of indirect blocks touched by mapping @nrblocks logically
686  * contiguous blocks
687  */
688 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
689 {
690         /*
691          * With N contiguous data blocks, we need at most
692          * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
693          * 2 dindirect blocks, and 1 tindirect block
694          */
695         return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
696 }
697
698 /*
699  * Truncate transactions can be complex and absolutely huge.  So we need to
700  * be able to restart the transaction at a conventient checkpoint to make
701  * sure we don't overflow the journal.
702  *
703  * Try to extend this transaction for the purposes of truncation.  If
704  * extend fails, we need to propagate the failure up and restart the
705  * transaction in the top-level truncate loop. --sct
706  *
707  * Returns 0 if we managed to create more room.  If we can't create more
708  * room, and the transaction must be restarted we return 1.
709  */
710 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
711 {
712         if (!ext4_handle_valid(handle))
713                 return 0;
714         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
715                 return 0;
716         if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
717                 return 0;
718         return 1;
719 }
720
721 /*
722  * Probably it should be a library function... search for first non-zero word
723  * or memcmp with zero_page, whatever is better for particular architecture.
724  * Linus?
725  */
726 static inline int all_zeroes(__le32 *p, __le32 *q)
727 {
728         while (p < q)
729                 if (*p++)
730                         return 0;
731         return 1;
732 }
733
734 /**
735  *      ext4_find_shared - find the indirect blocks for partial truncation.
736  *      @inode:   inode in question
737  *      @depth:   depth of the affected branch
738  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
739  *      @chain:   place to store the pointers to partial indirect blocks
740  *      @top:     place to the (detached) top of branch
741  *
742  *      This is a helper function used by ext4_truncate().
743  *
744  *      When we do truncate() we may have to clean the ends of several
745  *      indirect blocks but leave the blocks themselves alive. Block is
746  *      partially truncated if some data below the new i_size is referred
747  *      from it (and it is on the path to the first completely truncated
748  *      data block, indeed).  We have to free the top of that path along
749  *      with everything to the right of the path. Since no allocation
750  *      past the truncation point is possible until ext4_truncate()
751  *      finishes, we may safely do the latter, but top of branch may
752  *      require special attention - pageout below the truncation point
753  *      might try to populate it.
754  *
755  *      We atomically detach the top of branch from the tree, store the
756  *      block number of its root in *@top, pointers to buffer_heads of
757  *      partially truncated blocks - in @chain[].bh and pointers to
758  *      their last elements that should not be removed - in
759  *      @chain[].p. Return value is the pointer to last filled element
760  *      of @chain.
761  *
762  *      The work left to caller to do the actual freeing of subtrees:
763  *              a) free the subtree starting from *@top
764  *              b) free the subtrees whose roots are stored in
765  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
766  *              c) free the subtrees growing from the inode past the @chain[0].
767  *                      (no partially truncated stuff there).  */
768
769 static Indirect *ext4_find_shared(struct inode *inode, int depth,
770                                   ext4_lblk_t offsets[4], Indirect chain[4],
771                                   __le32 *top)
772 {
773         Indirect *partial, *p;
774         int k, err;
775
776         *top = 0;
777         /* Make k index the deepest non-null offset + 1 */
778         for (k = depth; k > 1 && !offsets[k-1]; k--)
779                 ;
780         partial = ext4_get_branch(inode, k, offsets, chain, &err);
781         /* Writer: pointers */
782         if (!partial)
783                 partial = chain + k-1;
784         /*
785          * If the branch acquired continuation since we've looked at it -
786          * fine, it should all survive and (new) top doesn't belong to us.
787          */
788         if (!partial->key && *partial->p)
789                 /* Writer: end */
790                 goto no_top;
791         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
792                 ;
793         /*
794          * OK, we've found the last block that must survive. The rest of our
795          * branch should be detached before unlocking. However, if that rest
796          * of branch is all ours and does not grow immediately from the inode
797          * it's easier to cheat and just decrement partial->p.
798          */
799         if (p == chain + k - 1 && p > chain) {
800                 p->p--;
801         } else {
802                 *top = *p->p;
803                 /* Nope, don't do this in ext4.  Must leave the tree intact */
804 #if 0
805                 *p->p = 0;
806 #endif
807         }
808         /* Writer: end */
809
810         while (partial > p) {
811                 brelse(partial->bh);
812                 partial--;
813         }
814 no_top:
815         return partial;
816 }
817
818 /*
819  * Zero a number of block pointers in either an inode or an indirect block.
820  * If we restart the transaction we must again get write access to the
821  * indirect block for further modification.
822  *
823  * We release `count' blocks on disk, but (last - first) may be greater
824  * than `count' because there can be holes in there.
825  *
826  * Return 0 on success, 1 on invalid block range
827  * and < 0 on fatal error.
828  */
829 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
830                              struct buffer_head *bh,
831                              ext4_fsblk_t block_to_free,
832                              unsigned long count, __le32 *first,
833                              __le32 *last)
834 {
835         __le32 *p;
836         int     flags = EXT4_FREE_BLOCKS_VALIDATED;
837         int     err;
838
839         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
840             ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
841                 flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
842         else if (ext4_should_journal_data(inode))
843                 flags |= EXT4_FREE_BLOCKS_FORGET;
844
845         if (!ext4_inode_block_valid(inode, block_to_free, count)) {
846                 EXT4_ERROR_INODE(inode, "attempt to clear invalid "
847                                  "blocks %llu len %lu",
848                                  (unsigned long long) block_to_free, count);
849                 return 1;
850         }
851
852         if (try_to_extend_transaction(handle, inode)) {
853                 if (bh) {
854                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
855                         err = ext4_handle_dirty_metadata(handle, inode, bh);
856                         if (unlikely(err))
857                                 goto out_err;
858                 }
859                 err = ext4_mark_inode_dirty(handle, inode);
860                 if (unlikely(err))
861                         goto out_err;
862                 err = ext4_truncate_restart_trans(handle, inode,
863                                         ext4_blocks_for_truncate(inode));
864                 if (unlikely(err))
865                         goto out_err;
866                 if (bh) {
867                         BUFFER_TRACE(bh, "retaking write access");
868                         err = ext4_journal_get_write_access(handle, bh);
869                         if (unlikely(err))
870                                 goto out_err;
871                 }
872         }
873
874         for (p = first; p < last; p++)
875                 *p = 0;
876
877         ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
878         return 0;
879 out_err:
880         ext4_std_error(inode->i_sb, err);
881         return err;
882 }
883
884 /**
885  * ext4_free_data - free a list of data blocks
886  * @handle:     handle for this transaction
887  * @inode:      inode we are dealing with
888  * @this_bh:    indirect buffer_head which contains *@first and *@last
889  * @first:      array of block numbers
890  * @last:       points immediately past the end of array
891  *
892  * We are freeing all blocks referred from that array (numbers are stored as
893  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
894  *
895  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
896  * blocks are contiguous then releasing them at one time will only affect one
897  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
898  * actually use a lot of journal space.
899  *
900  * @this_bh will be %NULL if @first and @last point into the inode's direct
901  * block pointers.
902  */
903 static void ext4_free_data(handle_t *handle, struct inode *inode,
904                            struct buffer_head *this_bh,
905                            __le32 *first, __le32 *last)
906 {
907         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
908         unsigned long count = 0;            /* Number of blocks in the run */
909         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
910                                                corresponding to
911                                                block_to_free */
912         ext4_fsblk_t nr;                    /* Current block # */
913         __le32 *p;                          /* Pointer into inode/ind
914                                                for current block */
915         int err = 0;
916
917         if (this_bh) {                          /* For indirect block */
918                 BUFFER_TRACE(this_bh, "get_write_access");
919                 err = ext4_journal_get_write_access(handle, this_bh);
920                 /* Important: if we can't update the indirect pointers
921                  * to the blocks, we can't free them. */
922                 if (err)
923                         return;
924         }
925
926         for (p = first; p < last; p++) {
927                 nr = le32_to_cpu(*p);
928                 if (nr) {
929                         /* accumulate blocks to free if they're contiguous */
930                         if (count == 0) {
931                                 block_to_free = nr;
932                                 block_to_free_p = p;
933                                 count = 1;
934                         } else if (nr == block_to_free + count) {
935                                 count++;
936                         } else {
937                                 err = ext4_clear_blocks(handle, inode, this_bh,
938                                                         block_to_free, count,
939                                                         block_to_free_p, p);
940                                 if (err)
941                                         break;
942                                 block_to_free = nr;
943                                 block_to_free_p = p;
944                                 count = 1;
945                         }
946                 }
947         }
948
949         if (!err && count > 0)
950                 err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
951                                         count, block_to_free_p, p);
952         if (err < 0)
953                 /* fatal error */
954                 return;
955
956         if (this_bh) {
957                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
958
959                 /*
960                  * The buffer head should have an attached journal head at this
961                  * point. However, if the data is corrupted and an indirect
962                  * block pointed to itself, it would have been detached when
963                  * the block was cleared. Check for this instead of OOPSing.
964                  */
965                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
966                         ext4_handle_dirty_metadata(handle, inode, this_bh);
967                 else
968                         EXT4_ERROR_INODE(inode,
969                                          "circular indirect block detected at "
970                                          "block %llu",
971                                 (unsigned long long) this_bh->b_blocknr);
972         }
973 }
974
975 /**
976  *      ext4_free_branches - free an array of branches
977  *      @handle: JBD handle for this transaction
978  *      @inode: inode we are dealing with
979  *      @parent_bh: the buffer_head which contains *@first and *@last
980  *      @first: array of block numbers
981  *      @last:  pointer immediately past the end of array
982  *      @depth: depth of the branches to free
983  *
984  *      We are freeing all blocks referred from these branches (numbers are
985  *      stored as little-endian 32-bit) and updating @inode->i_blocks
986  *      appropriately.
987  */
988 static void ext4_free_branches(handle_t *handle, struct inode *inode,
989                                struct buffer_head *parent_bh,
990                                __le32 *first, __le32 *last, int depth)
991 {
992         ext4_fsblk_t nr;
993         __le32 *p;
994
995         if (ext4_handle_is_aborted(handle))
996                 return;
997
998         if (depth--) {
999                 struct buffer_head *bh;
1000                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1001                 p = last;
1002                 while (--p >= first) {
1003                         nr = le32_to_cpu(*p);
1004                         if (!nr)
1005                                 continue;               /* A hole */
1006
1007                         if (!ext4_inode_block_valid(inode, nr, 1)) {
1008                                 EXT4_ERROR_INODE(inode,
1009                                                  "invalid indirect mapped "
1010                                                  "block %lu (level %d)",
1011                                                  (unsigned long) nr, depth);
1012                                 break;
1013                         }
1014
1015                         /* Go read the buffer for the next level down */
1016                         bh = sb_bread(inode->i_sb, nr);
1017
1018                         /*
1019                          * A read failure? Report error and clear slot
1020                          * (should be rare).
1021                          */
1022                         if (!bh) {
1023                                 EXT4_ERROR_INODE_BLOCK(inode, nr,
1024                                                        "Read failure");
1025                                 continue;
1026                         }
1027
1028                         /* This zaps the entire block.  Bottom up. */
1029                         BUFFER_TRACE(bh, "free child branches");
1030                         ext4_free_branches(handle, inode, bh,
1031                                         (__le32 *) bh->b_data,
1032                                         (__le32 *) bh->b_data + addr_per_block,
1033                                         depth);
1034                         brelse(bh);
1035
1036                         /*
1037                          * Everything below this this pointer has been
1038                          * released.  Now let this top-of-subtree go.
1039                          *
1040                          * We want the freeing of this indirect block to be
1041                          * atomic in the journal with the updating of the
1042                          * bitmap block which owns it.  So make some room in
1043                          * the journal.
1044                          *
1045                          * We zero the parent pointer *after* freeing its
1046                          * pointee in the bitmaps, so if extend_transaction()
1047                          * for some reason fails to put the bitmap changes and
1048                          * the release into the same transaction, recovery
1049                          * will merely complain about releasing a free block,
1050                          * rather than leaking blocks.
1051                          */
1052                         if (ext4_handle_is_aborted(handle))
1053                                 return;
1054                         if (try_to_extend_transaction(handle, inode)) {
1055                                 ext4_mark_inode_dirty(handle, inode);
1056                                 ext4_truncate_restart_trans(handle, inode,
1057                                             ext4_blocks_for_truncate(inode));
1058                         }
1059
1060                         /*
1061                          * The forget flag here is critical because if
1062                          * we are journaling (and not doing data
1063                          * journaling), we have to make sure a revoke
1064                          * record is written to prevent the journal
1065                          * replay from overwriting the (former)
1066                          * indirect block if it gets reallocated as a
1067                          * data block.  This must happen in the same
1068                          * transaction where the data blocks are
1069                          * actually freed.
1070                          */
1071                         ext4_free_blocks(handle, inode, NULL, nr, 1,
1072                                          EXT4_FREE_BLOCKS_METADATA|
1073                                          EXT4_FREE_BLOCKS_FORGET);
1074
1075                         if (parent_bh) {
1076                                 /*
1077                                  * The block which we have just freed is
1078                                  * pointed to by an indirect block: journal it
1079                                  */
1080                                 BUFFER_TRACE(parent_bh, "get_write_access");
1081                                 if (!ext4_journal_get_write_access(handle,
1082                                                                    parent_bh)){
1083                                         *p = 0;
1084                                         BUFFER_TRACE(parent_bh,
1085                                         "call ext4_handle_dirty_metadata");
1086                                         ext4_handle_dirty_metadata(handle,
1087                                                                    inode,
1088                                                                    parent_bh);
1089                                 }
1090                         }
1091                 }
1092         } else {
1093                 /* We have reached the bottom of the tree. */
1094                 BUFFER_TRACE(parent_bh, "free data blocks");
1095                 ext4_free_data(handle, inode, parent_bh, first, last);
1096         }
1097 }
1098
1099 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1100 {
1101         struct ext4_inode_info *ei = EXT4_I(inode);
1102         __le32 *i_data = ei->i_data;
1103         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1104         ext4_lblk_t offsets[4];
1105         Indirect chain[4];
1106         Indirect *partial;
1107         __le32 nr = 0;
1108         int n = 0;
1109         ext4_lblk_t last_block, max_block;
1110         unsigned blocksize = inode->i_sb->s_blocksize;
1111
1112         last_block = (inode->i_size + blocksize-1)
1113                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1114         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1115                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1116
1117         if (last_block != max_block) {
1118                 n = ext4_block_to_path(inode, last_block, offsets, NULL);
1119                 if (n == 0)
1120                         return;
1121         }
1122
1123         ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1124
1125         /*
1126          * The orphan list entry will now protect us from any crash which
1127          * occurs before the truncate completes, so it is now safe to propagate
1128          * the new, shorter inode size (held for now in i_size) into the
1129          * on-disk inode. We do this via i_disksize, which is the value which
1130          * ext4 *really* writes onto the disk inode.
1131          */
1132         ei->i_disksize = inode->i_size;
1133
1134         if (last_block == max_block) {
1135                 /*
1136                  * It is unnecessary to free any data blocks if last_block is
1137                  * equal to the indirect block limit.
1138                  */
1139                 return;
1140         } else if (n == 1) {            /* direct blocks */
1141                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1142                                i_data + EXT4_NDIR_BLOCKS);
1143                 goto do_indirects;
1144         }
1145
1146         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1147         /* Kill the top of shared branch (not detached) */
1148         if (nr) {
1149                 if (partial == chain) {
1150                         /* Shared branch grows from the inode */
1151                         ext4_free_branches(handle, inode, NULL,
1152                                            &nr, &nr+1, (chain+n-1) - partial);
1153                         *partial->p = 0;
1154                         /*
1155                          * We mark the inode dirty prior to restart,
1156                          * and prior to stop.  No need for it here.
1157                          */
1158                 } else {
1159                         /* Shared branch grows from an indirect block */
1160                         BUFFER_TRACE(partial->bh, "get_write_access");
1161                         ext4_free_branches(handle, inode, partial->bh,
1162                                         partial->p,
1163                                         partial->p+1, (chain+n-1) - partial);
1164                 }
1165         }
1166         /* Clear the ends of indirect blocks on the shared branch */
1167         while (partial > chain) {
1168                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1169                                    (__le32*)partial->bh->b_data+addr_per_block,
1170                                    (chain+n-1) - partial);
1171                 BUFFER_TRACE(partial->bh, "call brelse");
1172                 brelse(partial->bh);
1173                 partial--;
1174         }
1175 do_indirects:
1176         /* Kill the remaining (whole) subtrees */
1177         switch (offsets[0]) {
1178         default:
1179                 nr = i_data[EXT4_IND_BLOCK];
1180                 if (nr) {
1181                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1182                         i_data[EXT4_IND_BLOCK] = 0;
1183                 }
1184         case EXT4_IND_BLOCK:
1185                 nr = i_data[EXT4_DIND_BLOCK];
1186                 if (nr) {
1187                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1188                         i_data[EXT4_DIND_BLOCK] = 0;
1189                 }
1190         case EXT4_DIND_BLOCK:
1191                 nr = i_data[EXT4_TIND_BLOCK];
1192                 if (nr) {
1193                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1194                         i_data[EXT4_TIND_BLOCK] = 0;
1195                 }
1196         case EXT4_TIND_BLOCK:
1197                 ;
1198         }
1199 }
1200
1201 /**
1202  *      ext4_ind_remove_space - remove space from the range
1203  *      @handle: JBD handle for this transaction
1204  *      @inode: inode we are dealing with
1205  *      @start: First block to remove
1206  *      @end:   One block after the last block to remove (exclusive)
1207  *
1208  *      Free the blocks in the defined range (end is exclusive endpoint of
1209  *      range). This is used by ext4_punch_hole().
1210  */
1211 int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1212                           ext4_lblk_t start, ext4_lblk_t end)
1213 {
1214         struct ext4_inode_info *ei = EXT4_I(inode);
1215         __le32 *i_data = ei->i_data;
1216         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1217         ext4_lblk_t offsets[4], offsets2[4];
1218         Indirect chain[4], chain2[4];
1219         Indirect *partial, *partial2;
1220         Indirect *p = NULL, *p2 = NULL;
1221         ext4_lblk_t max_block;
1222         __le32 nr = 0, nr2 = 0;
1223         int n = 0, n2 = 0;
1224         unsigned blocksize = inode->i_sb->s_blocksize;
1225
1226         max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1227                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1228         if (end >= max_block)
1229                 end = max_block;
1230         if ((start >= end) || (start > max_block))
1231                 return 0;
1232
1233         n = ext4_block_to_path(inode, start, offsets, NULL);
1234         n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1235
1236         BUG_ON(n > n2);
1237
1238         if ((n == 1) && (n == n2)) {
1239                 /* We're punching only within direct block range */
1240                 ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1241                                i_data + offsets2[0]);
1242                 return 0;
1243         } else if (n2 > n) {
1244                 /*
1245                  * Start and end are on a different levels so we're going to
1246                  * free partial block at start, and partial block at end of
1247                  * the range. If there are some levels in between then
1248                  * do_indirects label will take care of that.
1249                  */
1250
1251                 if (n == 1) {
1252                         /*
1253                          * Start is at the direct block level, free
1254                          * everything to the end of the level.
1255                          */
1256                         ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1257                                        i_data + EXT4_NDIR_BLOCKS);
1258                         goto end_range;
1259                 }
1260
1261
1262                 partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1263                 if (nr) {
1264                         if (partial == chain) {
1265                                 /* Shared branch grows from the inode */
1266                                 ext4_free_branches(handle, inode, NULL,
1267                                            &nr, &nr+1, (chain+n-1) - partial);
1268                                 *partial->p = 0;
1269                         } else {
1270                                 /* Shared branch grows from an indirect block */
1271                                 BUFFER_TRACE(partial->bh, "get_write_access");
1272                                 ext4_free_branches(handle, inode, partial->bh,
1273                                         partial->p,
1274                                         partial->p+1, (chain+n-1) - partial);
1275                         }
1276                 }
1277
1278                 /*
1279                  * Clear the ends of indirect blocks on the shared branch
1280                  * at the start of the range
1281                  */
1282                 while (partial > chain) {
1283                         ext4_free_branches(handle, inode, partial->bh,
1284                                 partial->p + 1,
1285                                 (__le32 *)partial->bh->b_data+addr_per_block,
1286                                 (chain+n-1) - partial);
1287                         partial--;
1288                 }
1289
1290 end_range:
1291                 partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1292                 if (nr2) {
1293                         if (partial2 == chain2) {
1294                                 /*
1295                                  * Remember, end is exclusive so here we're at
1296                                  * the start of the next level we're not going
1297                                  * to free. Everything was covered by the start
1298                                  * of the range.
1299                                  */
1300                                 goto do_indirects;
1301                         }
1302                 } else {
1303                         /*
1304                          * ext4_find_shared returns Indirect structure which
1305                          * points to the last element which should not be
1306                          * removed by truncate. But this is end of the range
1307                          * in punch_hole so we need to point to the next element
1308                          */
1309                         partial2->p++;
1310                 }
1311
1312                 /*
1313                  * Clear the ends of indirect blocks on the shared branch
1314                  * at the end of the range
1315                  */
1316                 while (partial2 > chain2) {
1317                         ext4_free_branches(handle, inode, partial2->bh,
1318                                            (__le32 *)partial2->bh->b_data,
1319                                            partial2->p,
1320                                            (chain2+n2-1) - partial2);
1321                         partial2--;
1322                 }
1323                 goto do_indirects;
1324         }
1325
1326         /* Punch happened within the same level (n == n2) */
1327         partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1328         partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1329
1330         /* Free top, but only if partial2 isn't its subtree. */
1331         if (nr) {
1332                 int level = min(partial - chain, partial2 - chain2);
1333                 int i;
1334                 int subtree = 1;
1335
1336                 for (i = 0; i <= level; i++) {
1337                         if (offsets[i] != offsets2[i]) {
1338                                 subtree = 0;
1339                                 break;
1340                         }
1341                 }
1342
1343                 if (!subtree) {
1344                         if (partial == chain) {
1345                                 /* Shared branch grows from the inode */
1346                                 ext4_free_branches(handle, inode, NULL,
1347                                                    &nr, &nr+1,
1348                                                    (chain+n-1) - partial);
1349                                 *partial->p = 0;
1350                         } else {
1351                                 /* Shared branch grows from an indirect block */
1352                                 BUFFER_TRACE(partial->bh, "get_write_access");
1353                                 ext4_free_branches(handle, inode, partial->bh,
1354                                                    partial->p,
1355                                                    partial->p+1,
1356                                                    (chain+n-1) - partial);
1357                         }
1358                 }
1359         }
1360
1361         if (!nr2) {
1362                 /*
1363                  * ext4_find_shared returns Indirect structure which
1364                  * points to the last element which should not be
1365                  * removed by truncate. But this is end of the range
1366                  * in punch_hole so we need to point to the next element
1367                  */
1368                 partial2->p++;
1369         }
1370
1371         while (partial > chain || partial2 > chain2) {
1372                 int depth = (chain+n-1) - partial;
1373                 int depth2 = (chain2+n2-1) - partial2;
1374
1375                 if (partial > chain && partial2 > chain2 &&
1376                     partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1377                         /*
1378                          * We've converged on the same block. Clear the range,
1379                          * then we're done.
1380                          */
1381                         ext4_free_branches(handle, inode, partial->bh,
1382                                            partial->p + 1,
1383                                            partial2->p,
1384                                            (chain+n-1) - partial);
1385                         goto cleanup;
1386                 }
1387
1388                 /*
1389                  * The start and end partial branches may not be at the same
1390                  * level even though the punch happened within one level. So, we
1391                  * give them a chance to arrive at the same level, then walk
1392                  * them in step with each other until we converge on the same
1393                  * block.
1394                  */
1395                 if (partial > chain && depth <= depth2) {
1396                         ext4_free_branches(handle, inode, partial->bh,
1397                                            partial->p + 1,
1398                                            (__le32 *)partial->bh->b_data+addr_per_block,
1399                                            (chain+n-1) - partial);
1400                         partial--;
1401                 }
1402                 if (partial2 > chain2 && depth2 <= depth) {
1403                         ext4_free_branches(handle, inode, partial2->bh,
1404                                            (__le32 *)partial2->bh->b_data,
1405                                            partial2->p,
1406                                            (chain2+n2-1) - partial2);
1407                         partial2--;
1408                 }
1409         }
1410
1411 cleanup:
1412         while (p && p > chain) {
1413                 BUFFER_TRACE(p->bh, "call brelse");
1414                 brelse(p->bh);
1415                 p--;
1416         }
1417         while (p2 && p2 > chain2) {
1418                 BUFFER_TRACE(p2->bh, "call brelse");
1419                 brelse(p2->bh);
1420                 p2--;
1421         }
1422         return 0;
1423
1424 do_indirects:
1425         /* Kill the remaining (whole) subtrees */
1426         switch (offsets[0]) {
1427         default:
1428                 if (++n >= n2)
1429                         break;
1430                 nr = i_data[EXT4_IND_BLOCK];
1431                 if (nr) {
1432                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1433                         i_data[EXT4_IND_BLOCK] = 0;
1434                 }
1435         case EXT4_IND_BLOCK:
1436                 if (++n >= n2)
1437                         break;
1438                 nr = i_data[EXT4_DIND_BLOCK];
1439                 if (nr) {
1440                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1441                         i_data[EXT4_DIND_BLOCK] = 0;
1442                 }
1443         case EXT4_DIND_BLOCK:
1444                 if (++n >= n2)
1445                         break;
1446                 nr = i_data[EXT4_TIND_BLOCK];
1447                 if (nr) {
1448                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1449                         i_data[EXT4_TIND_BLOCK] = 0;
1450                 }
1451         case EXT4_TIND_BLOCK:
1452                 ;
1453         }
1454         goto cleanup;
1455 }