xref: /openbmc/linux/fs/ext4/indirect.c (revision b8d312aa)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/ext4/indirect.c
4  *
5  *  from
6  *
7  *  linux/fs/ext4/inode.c
8  *
9  * Copyright (C) 1992, 1993, 1994, 1995
10  * Remy Card (card@masi.ibp.fr)
11  * Laboratoire MASI - Institut Blaise Pascal
12  * Universite Pierre et Marie Curie (Paris VI)
13  *
14  *  from
15  *
16  *  linux/fs/minix/inode.c
17  *
18  *  Copyright (C) 1991, 1992  Linus Torvalds
19  *
20  *  Goal-directed block allocation by Stephen Tweedie
21  *	(sct@redhat.com), 1993, 1998
22  */
23 
24 #include "ext4_jbd2.h"
25 #include "truncate.h"
26 #include <linux/dax.h>
27 #include <linux/uio.h>
28 
29 #include <trace/events/ext4.h>
30 
31 typedef struct {
32 	__le32	*p;
33 	__le32	key;
34 	struct buffer_head *bh;
35 } Indirect;
36 
37 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
38 {
39 	p->key = *(p->p = v);
40 	p->bh = bh;
41 }
42 
43 /**
44  *	ext4_block_to_path - parse the block number into array of offsets
45  *	@inode: inode in question (we are only interested in its superblock)
46  *	@i_block: block number to be parsed
47  *	@offsets: array to store the offsets in
48  *	@boundary: set this non-zero if the referred-to block is likely to be
49  *	       followed (on disk) by an indirect block.
50  *
51  *	To store the locations of file's data ext4 uses a data structure common
52  *	for UNIX filesystems - tree of pointers anchored in the inode, with
53  *	data blocks at leaves and indirect blocks in intermediate nodes.
54  *	This function translates the block number into path in that tree -
55  *	return value is the path length and @offsets[n] is the offset of
56  *	pointer to (n+1)th node in the nth one. If @block is out of range
57  *	(negative or too large) warning is printed and zero returned.
58  *
59  *	Note: function doesn't find node addresses, so no IO is needed. All
60  *	we need to know is the capacity of indirect blocks (taken from the
61  *	inode->i_sb).
62  */
63 
64 /*
65  * Portability note: the last comparison (check that we fit into triple
66  * indirect block) is spelled differently, because otherwise on an
67  * architecture with 32-bit longs and 8Kb pages we might get into trouble
68  * if our filesystem had 8Kb blocks. We might use long long, but that would
69  * kill us on x86. Oh, well, at least the sign propagation does not matter -
70  * i_block would have to be negative in the very beginning, so we would not
71  * get there at all.
72  */
73 
74 static int ext4_block_to_path(struct inode *inode,
75 			      ext4_lblk_t i_block,
76 			      ext4_lblk_t offsets[4], int *boundary)
77 {
78 	int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
79 	int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
80 	const long direct_blocks = EXT4_NDIR_BLOCKS,
81 		indirect_blocks = ptrs,
82 		double_blocks = (1 << (ptrs_bits * 2));
83 	int n = 0;
84 	int final = 0;
85 
86 	if (i_block < direct_blocks) {
87 		offsets[n++] = i_block;
88 		final = direct_blocks;
89 	} else if ((i_block -= direct_blocks) < indirect_blocks) {
90 		offsets[n++] = EXT4_IND_BLOCK;
91 		offsets[n++] = i_block;
92 		final = ptrs;
93 	} else if ((i_block -= indirect_blocks) < double_blocks) {
94 		offsets[n++] = EXT4_DIND_BLOCK;
95 		offsets[n++] = i_block >> ptrs_bits;
96 		offsets[n++] = i_block & (ptrs - 1);
97 		final = ptrs;
98 	} else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
99 		offsets[n++] = EXT4_TIND_BLOCK;
100 		offsets[n++] = i_block >> (ptrs_bits * 2);
101 		offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
102 		offsets[n++] = i_block & (ptrs - 1);
103 		final = ptrs;
104 	} else {
105 		ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
106 			     i_block + direct_blocks +
107 			     indirect_blocks + double_blocks, inode->i_ino);
108 	}
109 	if (boundary)
110 		*boundary = final - 1 - (i_block & (ptrs - 1));
111 	return n;
112 }
113 
114 /**
115  *	ext4_get_branch - read the chain of indirect blocks leading to data
116  *	@inode: inode in question
117  *	@depth: depth of the chain (1 - direct pointer, etc.)
118  *	@offsets: offsets of pointers in inode/indirect blocks
119  *	@chain: place to store the result
120  *	@err: here we store the error value
121  *
122  *	Function fills the array of triples <key, p, bh> and returns %NULL
123  *	if everything went OK or the pointer to the last filled triple
124  *	(incomplete one) otherwise. Upon the return chain[i].key contains
125  *	the number of (i+1)-th block in the chain (as it is stored in memory,
126  *	i.e. little-endian 32-bit), chain[i].p contains the address of that
127  *	number (it points into struct inode for i==0 and into the bh->b_data
128  *	for i>0) and chain[i].bh points to the buffer_head of i-th indirect
129  *	block for i>0 and NULL for i==0. In other words, it holds the block
130  *	numbers of the chain, addresses they were taken from (and where we can
131  *	verify that chain did not change) and buffer_heads hosting these
132  *	numbers.
133  *
134  *	Function stops when it stumbles upon zero pointer (absent block)
135  *		(pointer to last triple returned, *@err == 0)
136  *	or when it gets an IO error reading an indirect block
137  *		(ditto, *@err == -EIO)
138  *	or when it reads all @depth-1 indirect blocks successfully and finds
139  *	the whole chain, all way to the data (returns %NULL, *err == 0).
140  *
141  *      Need to be called with
142  *      down_read(&EXT4_I(inode)->i_data_sem)
143  */
144 static Indirect *ext4_get_branch(struct inode *inode, int depth,
145 				 ext4_lblk_t  *offsets,
146 				 Indirect chain[4], int *err)
147 {
148 	struct super_block *sb = inode->i_sb;
149 	Indirect *p = chain;
150 	struct buffer_head *bh;
151 	int ret = -EIO;
152 
153 	*err = 0;
154 	/* i_data is not going away, no lock needed */
155 	add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
156 	if (!p->key)
157 		goto no_block;
158 	while (--depth) {
159 		bh = sb_getblk(sb, le32_to_cpu(p->key));
160 		if (unlikely(!bh)) {
161 			ret = -ENOMEM;
162 			goto failure;
163 		}
164 
165 		if (!bh_uptodate_or_lock(bh)) {
166 			if (bh_submit_read(bh) < 0) {
167 				put_bh(bh);
168 				goto failure;
169 			}
170 			/* validate block references */
171 			if (ext4_check_indirect_blockref(inode, bh)) {
172 				put_bh(bh);
173 				goto failure;
174 			}
175 		}
176 
177 		add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
178 		/* Reader: end */
179 		if (!p->key)
180 			goto no_block;
181 	}
182 	return NULL;
183 
184 failure:
185 	*err = ret;
186 no_block:
187 	return p;
188 }
189 
190 /**
191  *	ext4_find_near - find a place for allocation with sufficient locality
192  *	@inode: owner
193  *	@ind: descriptor of indirect block.
194  *
195  *	This function returns the preferred place for block allocation.
196  *	It is used when heuristic for sequential allocation fails.
197  *	Rules are:
198  *	  + if there is a block to the left of our position - allocate near it.
199  *	  + if pointer will live in indirect block - allocate near that block.
200  *	  + if pointer will live in inode - allocate in the same
201  *	    cylinder group.
202  *
203  * In the latter case we colour the starting block by the callers PID to
204  * prevent it from clashing with concurrent allocations for a different inode
205  * in the same block group.   The PID is used here so that functionally related
206  * files will be close-by on-disk.
207  *
208  *	Caller must make sure that @ind is valid and will stay that way.
209  */
210 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
211 {
212 	struct ext4_inode_info *ei = EXT4_I(inode);
213 	__le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
214 	__le32 *p;
215 
216 	/* Try to find previous block */
217 	for (p = ind->p - 1; p >= start; p--) {
218 		if (*p)
219 			return le32_to_cpu(*p);
220 	}
221 
222 	/* No such thing, so let's try location of indirect block */
223 	if (ind->bh)
224 		return ind->bh->b_blocknr;
225 
226 	/*
227 	 * It is going to be referred to from the inode itself? OK, just put it
228 	 * into the same cylinder group then.
229 	 */
230 	return ext4_inode_to_goal_block(inode);
231 }
232 
233 /**
234  *	ext4_find_goal - find a preferred place for allocation.
235  *	@inode: owner
236  *	@block:  block we want
237  *	@partial: pointer to the last triple within a chain
238  *
239  *	Normally this function find the preferred place for block allocation,
240  *	returns it.
241  *	Because this is only used for non-extent files, we limit the block nr
242  *	to 32 bits.
243  */
244 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
245 				   Indirect *partial)
246 {
247 	ext4_fsblk_t goal;
248 
249 	/*
250 	 * XXX need to get goal block from mballoc's data structures
251 	 */
252 
253 	goal = ext4_find_near(inode, partial);
254 	goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
255 	return goal;
256 }
257 
258 /**
259  *	ext4_blks_to_allocate - Look up the block map and count the number
260  *	of direct blocks need to be allocated for the given branch.
261  *
262  *	@branch: chain of indirect blocks
263  *	@k: number of blocks need for indirect blocks
264  *	@blks: number of data blocks to be mapped.
265  *	@blocks_to_boundary:  the offset in the indirect block
266  *
267  *	return the total number of blocks to be allocate, including the
268  *	direct and indirect blocks.
269  */
270 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
271 				 int blocks_to_boundary)
272 {
273 	unsigned int count = 0;
274 
275 	/*
276 	 * Simple case, [t,d]Indirect block(s) has not allocated yet
277 	 * then it's clear blocks on that path have not allocated
278 	 */
279 	if (k > 0) {
280 		/* right now we don't handle cross boundary allocation */
281 		if (blks < blocks_to_boundary + 1)
282 			count += blks;
283 		else
284 			count += blocks_to_boundary + 1;
285 		return count;
286 	}
287 
288 	count++;
289 	while (count < blks && count <= blocks_to_boundary &&
290 		le32_to_cpu(*(branch[0].p + count)) == 0) {
291 		count++;
292 	}
293 	return count;
294 }
295 
296 /**
297  * ext4_alloc_branch() - allocate and set up a chain of blocks
298  * @handle: handle for this transaction
299  * @ar: structure describing the allocation request
300  * @indirect_blks: number of allocated indirect blocks
301  * @offsets: offsets (in the blocks) to store the pointers to next.
302  * @branch: place to store the chain in.
303  *
304  *	This function allocates blocks, zeroes out all but the last one,
305  *	links them into chain and (if we are synchronous) writes them to disk.
306  *	In other words, it prepares a branch that can be spliced onto the
307  *	inode. It stores the information about that chain in the branch[], in
308  *	the same format as ext4_get_branch() would do. We are calling it after
309  *	we had read the existing part of chain and partial points to the last
310  *	triple of that (one with zero ->key). Upon the exit we have the same
311  *	picture as after the successful ext4_get_block(), except that in one
312  *	place chain is disconnected - *branch->p is still zero (we did not
313  *	set the last link), but branch->key contains the number that should
314  *	be placed into *branch->p to fill that gap.
315  *
316  *	If allocation fails we free all blocks we've allocated (and forget
317  *	their buffer_heads) and return the error value the from failed
318  *	ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
319  *	as described above and return 0.
320  */
321 static int ext4_alloc_branch(handle_t *handle,
322 			     struct ext4_allocation_request *ar,
323 			     int indirect_blks, ext4_lblk_t *offsets,
324 			     Indirect *branch)
325 {
326 	struct buffer_head *		bh;
327 	ext4_fsblk_t			b, new_blocks[4];
328 	__le32				*p;
329 	int				i, j, err, len = 1;
330 
331 	for (i = 0; i <= indirect_blks; i++) {
332 		if (i == indirect_blks) {
333 			new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
334 		} else
335 			ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
336 					ar->inode, ar->goal,
337 					ar->flags & EXT4_MB_DELALLOC_RESERVED,
338 					NULL, &err);
339 		if (err) {
340 			i--;
341 			goto failed;
342 		}
343 		branch[i].key = cpu_to_le32(new_blocks[i]);
344 		if (i == 0)
345 			continue;
346 
347 		bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
348 		if (unlikely(!bh)) {
349 			err = -ENOMEM;
350 			goto failed;
351 		}
352 		lock_buffer(bh);
353 		BUFFER_TRACE(bh, "call get_create_access");
354 		err = ext4_journal_get_create_access(handle, bh);
355 		if (err) {
356 			unlock_buffer(bh);
357 			goto failed;
358 		}
359 
360 		memset(bh->b_data, 0, bh->b_size);
361 		p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
362 		b = new_blocks[i];
363 
364 		if (i == indirect_blks)
365 			len = ar->len;
366 		for (j = 0; j < len; j++)
367 			*p++ = cpu_to_le32(b++);
368 
369 		BUFFER_TRACE(bh, "marking uptodate");
370 		set_buffer_uptodate(bh);
371 		unlock_buffer(bh);
372 
373 		BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
374 		err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
375 		if (err)
376 			goto failed;
377 	}
378 	return 0;
379 failed:
380 	for (; i >= 0; i--) {
381 		/*
382 		 * We want to ext4_forget() only freshly allocated indirect
383 		 * blocks.  Buffer for new_blocks[i-1] is at branch[i].bh and
384 		 * buffer at branch[0].bh is indirect block / inode already
385 		 * existing before ext4_alloc_branch() was called.
386 		 */
387 		if (i > 0 && i != indirect_blks && branch[i].bh)
388 			ext4_forget(handle, 1, ar->inode, branch[i].bh,
389 				    branch[i].bh->b_blocknr);
390 		ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
391 				 (i == indirect_blks) ? ar->len : 1, 0);
392 	}
393 	return err;
394 }
395 
396 /**
397  * ext4_splice_branch() - splice the allocated branch onto inode.
398  * @handle: handle for this transaction
399  * @ar: structure describing the allocation request
400  * @where: location of missing link
401  * @num:   number of indirect blocks we are adding
402  *
403  * This function fills the missing link and does all housekeeping needed in
404  * inode (->i_blocks, etc.). In case of success we end up with the full
405  * chain to new block and return 0.
406  */
407 static int ext4_splice_branch(handle_t *handle,
408 			      struct ext4_allocation_request *ar,
409 			      Indirect *where, int num)
410 {
411 	int i;
412 	int err = 0;
413 	ext4_fsblk_t current_block;
414 
415 	/*
416 	 * If we're splicing into a [td]indirect block (as opposed to the
417 	 * inode) then we need to get write access to the [td]indirect block
418 	 * before the splice.
419 	 */
420 	if (where->bh) {
421 		BUFFER_TRACE(where->bh, "get_write_access");
422 		err = ext4_journal_get_write_access(handle, where->bh);
423 		if (err)
424 			goto err_out;
425 	}
426 	/* That's it */
427 
428 	*where->p = where->key;
429 
430 	/*
431 	 * Update the host buffer_head or inode to point to more just allocated
432 	 * direct blocks blocks
433 	 */
434 	if (num == 0 && ar->len > 1) {
435 		current_block = le32_to_cpu(where->key) + 1;
436 		for (i = 1; i < ar->len; i++)
437 			*(where->p + i) = cpu_to_le32(current_block++);
438 	}
439 
440 	/* We are done with atomic stuff, now do the rest of housekeeping */
441 	/* had we spliced it onto indirect block? */
442 	if (where->bh) {
443 		/*
444 		 * If we spliced it onto an indirect block, we haven't
445 		 * altered the inode.  Note however that if it is being spliced
446 		 * onto an indirect block at the very end of the file (the
447 		 * file is growing) then we *will* alter the inode to reflect
448 		 * the new i_size.  But that is not done here - it is done in
449 		 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
450 		 */
451 		jbd_debug(5, "splicing indirect only\n");
452 		BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
453 		err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
454 		if (err)
455 			goto err_out;
456 	} else {
457 		/*
458 		 * OK, we spliced it into the inode itself on a direct block.
459 		 */
460 		ext4_mark_inode_dirty(handle, ar->inode);
461 		jbd_debug(5, "splicing direct\n");
462 	}
463 	return err;
464 
465 err_out:
466 	for (i = 1; i <= num; i++) {
467 		/*
468 		 * branch[i].bh is newly allocated, so there is no
469 		 * need to revoke the block, which is why we don't
470 		 * need to set EXT4_FREE_BLOCKS_METADATA.
471 		 */
472 		ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
473 				 EXT4_FREE_BLOCKS_FORGET);
474 	}
475 	ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
476 			 ar->len, 0);
477 
478 	return err;
479 }
480 
481 /*
482  * The ext4_ind_map_blocks() function handles non-extents inodes
483  * (i.e., using the traditional indirect/double-indirect i_blocks
484  * scheme) for ext4_map_blocks().
485  *
486  * Allocation strategy is simple: if we have to allocate something, we will
487  * have to go the whole way to leaf. So let's do it before attaching anything
488  * to tree, set linkage between the newborn blocks, write them if sync is
489  * required, recheck the path, free and repeat if check fails, otherwise
490  * set the last missing link (that will protect us from any truncate-generated
491  * removals - all blocks on the path are immune now) and possibly force the
492  * write on the parent block.
493  * That has a nice additional property: no special recovery from the failed
494  * allocations is needed - we simply release blocks and do not touch anything
495  * reachable from inode.
496  *
497  * `handle' can be NULL if create == 0.
498  *
499  * return > 0, # of blocks mapped or allocated.
500  * return = 0, if plain lookup failed.
501  * return < 0, error case.
502  *
503  * The ext4_ind_get_blocks() function should be called with
504  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
505  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
506  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
507  * blocks.
508  */
509 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
510 			struct ext4_map_blocks *map,
511 			int flags)
512 {
513 	struct ext4_allocation_request ar;
514 	int err = -EIO;
515 	ext4_lblk_t offsets[4];
516 	Indirect chain[4];
517 	Indirect *partial;
518 	int indirect_blks;
519 	int blocks_to_boundary = 0;
520 	int depth;
521 	int count = 0;
522 	ext4_fsblk_t first_block = 0;
523 
524 	trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
525 	J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
526 	J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
527 	depth = ext4_block_to_path(inode, map->m_lblk, offsets,
528 				   &blocks_to_boundary);
529 
530 	if (depth == 0)
531 		goto out;
532 
533 	partial = ext4_get_branch(inode, depth, offsets, chain, &err);
534 
535 	/* Simplest case - block found, no allocation needed */
536 	if (!partial) {
537 		first_block = le32_to_cpu(chain[depth - 1].key);
538 		count++;
539 		/*map more blocks*/
540 		while (count < map->m_len && count <= blocks_to_boundary) {
541 			ext4_fsblk_t blk;
542 
543 			blk = le32_to_cpu(*(chain[depth-1].p + count));
544 
545 			if (blk == first_block + count)
546 				count++;
547 			else
548 				break;
549 		}
550 		goto got_it;
551 	}
552 
553 	/* Next simple case - plain lookup failed */
554 	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
555 		unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
556 		int i;
557 
558 		/*
559 		 * Count number blocks in a subtree under 'partial'. At each
560 		 * level we count number of complete empty subtrees beyond
561 		 * current offset and then descend into the subtree only
562 		 * partially beyond current offset.
563 		 */
564 		count = 0;
565 		for (i = partial - chain + 1; i < depth; i++)
566 			count = count * epb + (epb - offsets[i] - 1);
567 		count++;
568 		/* Fill in size of a hole we found */
569 		map->m_pblk = 0;
570 		map->m_len = min_t(unsigned int, map->m_len, count);
571 		goto cleanup;
572 	}
573 
574 	/* Failed read of indirect block */
575 	if (err == -EIO)
576 		goto cleanup;
577 
578 	/*
579 	 * Okay, we need to do block allocation.
580 	*/
581 	if (ext4_has_feature_bigalloc(inode->i_sb)) {
582 		EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
583 				 "non-extent mapped inodes with bigalloc");
584 		return -EFSCORRUPTED;
585 	}
586 
587 	/* Set up for the direct block allocation */
588 	memset(&ar, 0, sizeof(ar));
589 	ar.inode = inode;
590 	ar.logical = map->m_lblk;
591 	if (S_ISREG(inode->i_mode))
592 		ar.flags = EXT4_MB_HINT_DATA;
593 	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
594 		ar.flags |= EXT4_MB_DELALLOC_RESERVED;
595 	if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
596 		ar.flags |= EXT4_MB_USE_RESERVED;
597 
598 	ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
599 
600 	/* the number of blocks need to allocate for [d,t]indirect blocks */
601 	indirect_blks = (chain + depth) - partial - 1;
602 
603 	/*
604 	 * Next look up the indirect map to count the totoal number of
605 	 * direct blocks to allocate for this branch.
606 	 */
607 	ar.len = ext4_blks_to_allocate(partial, indirect_blks,
608 				       map->m_len, blocks_to_boundary);
609 
610 	/*
611 	 * Block out ext4_truncate while we alter the tree
612 	 */
613 	err = ext4_alloc_branch(handle, &ar, indirect_blks,
614 				offsets + (partial - chain), partial);
615 
616 	/*
617 	 * The ext4_splice_branch call will free and forget any buffers
618 	 * on the new chain if there is a failure, but that risks using
619 	 * up transaction credits, especially for bitmaps where the
620 	 * credits cannot be returned.  Can we handle this somehow?  We
621 	 * may need to return -EAGAIN upwards in the worst case.  --sct
622 	 */
623 	if (!err)
624 		err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
625 	if (err)
626 		goto cleanup;
627 
628 	map->m_flags |= EXT4_MAP_NEW;
629 
630 	ext4_update_inode_fsync_trans(handle, inode, 1);
631 	count = ar.len;
632 got_it:
633 	map->m_flags |= EXT4_MAP_MAPPED;
634 	map->m_pblk = le32_to_cpu(chain[depth-1].key);
635 	map->m_len = count;
636 	if (count > blocks_to_boundary)
637 		map->m_flags |= EXT4_MAP_BOUNDARY;
638 	err = count;
639 	/* Clean up and exit */
640 	partial = chain + depth - 1;	/* the whole chain */
641 cleanup:
642 	while (partial > chain) {
643 		BUFFER_TRACE(partial->bh, "call brelse");
644 		brelse(partial->bh);
645 		partial--;
646 	}
647 out:
648 	trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
649 	return err;
650 }
651 
652 /*
653  * Calculate the number of metadata blocks need to reserve
654  * to allocate a new block at @lblocks for non extent file based file
655  */
656 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
657 {
658 	struct ext4_inode_info *ei = EXT4_I(inode);
659 	sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
660 	int blk_bits;
661 
662 	if (lblock < EXT4_NDIR_BLOCKS)
663 		return 0;
664 
665 	lblock -= EXT4_NDIR_BLOCKS;
666 
667 	if (ei->i_da_metadata_calc_len &&
668 	    (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
669 		ei->i_da_metadata_calc_len++;
670 		return 0;
671 	}
672 	ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
673 	ei->i_da_metadata_calc_len = 1;
674 	blk_bits = order_base_2(lblock);
675 	return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
676 }
677 
678 /*
679  * Calculate number of indirect blocks touched by mapping @nrblocks logically
680  * contiguous blocks
681  */
682 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
683 {
684 	/*
685 	 * With N contiguous data blocks, we need at most
686 	 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
687 	 * 2 dindirect blocks, and 1 tindirect block
688 	 */
689 	return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
690 }
691 
692 /*
693  * Truncate transactions can be complex and absolutely huge.  So we need to
694  * be able to restart the transaction at a conventient checkpoint to make
695  * sure we don't overflow the journal.
696  *
697  * Try to extend this transaction for the purposes of truncation.  If
698  * extend fails, we need to propagate the failure up and restart the
699  * transaction in the top-level truncate loop. --sct
700  *
701  * Returns 0 if we managed to create more room.  If we can't create more
702  * room, and the transaction must be restarted we return 1.
703  */
704 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
705 {
706 	if (!ext4_handle_valid(handle))
707 		return 0;
708 	if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
709 		return 0;
710 	if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
711 		return 0;
712 	return 1;
713 }
714 
715 /*
716  * Probably it should be a library function... search for first non-zero word
717  * or memcmp with zero_page, whatever is better for particular architecture.
718  * Linus?
719  */
720 static inline int all_zeroes(__le32 *p, __le32 *q)
721 {
722 	while (p < q)
723 		if (*p++)
724 			return 0;
725 	return 1;
726 }
727 
728 /**
729  *	ext4_find_shared - find the indirect blocks for partial truncation.
730  *	@inode:	  inode in question
731  *	@depth:	  depth of the affected branch
732  *	@offsets: offsets of pointers in that branch (see ext4_block_to_path)
733  *	@chain:	  place to store the pointers to partial indirect blocks
734  *	@top:	  place to the (detached) top of branch
735  *
736  *	This is a helper function used by ext4_truncate().
737  *
738  *	When we do truncate() we may have to clean the ends of several
739  *	indirect blocks but leave the blocks themselves alive. Block is
740  *	partially truncated if some data below the new i_size is referred
741  *	from it (and it is on the path to the first completely truncated
742  *	data block, indeed).  We have to free the top of that path along
743  *	with everything to the right of the path. Since no allocation
744  *	past the truncation point is possible until ext4_truncate()
745  *	finishes, we may safely do the latter, but top of branch may
746  *	require special attention - pageout below the truncation point
747  *	might try to populate it.
748  *
749  *	We atomically detach the top of branch from the tree, store the
750  *	block number of its root in *@top, pointers to buffer_heads of
751  *	partially truncated blocks - in @chain[].bh and pointers to
752  *	their last elements that should not be removed - in
753  *	@chain[].p. Return value is the pointer to last filled element
754  *	of @chain.
755  *
756  *	The work left to caller to do the actual freeing of subtrees:
757  *		a) free the subtree starting from *@top
758  *		b) free the subtrees whose roots are stored in
759  *			(@chain[i].p+1 .. end of @chain[i].bh->b_data)
760  *		c) free the subtrees growing from the inode past the @chain[0].
761  *			(no partially truncated stuff there).  */
762 
763 static Indirect *ext4_find_shared(struct inode *inode, int depth,
764 				  ext4_lblk_t offsets[4], Indirect chain[4],
765 				  __le32 *top)
766 {
767 	Indirect *partial, *p;
768 	int k, err;
769 
770 	*top = 0;
771 	/* Make k index the deepest non-null offset + 1 */
772 	for (k = depth; k > 1 && !offsets[k-1]; k--)
773 		;
774 	partial = ext4_get_branch(inode, k, offsets, chain, &err);
775 	/* Writer: pointers */
776 	if (!partial)
777 		partial = chain + k-1;
778 	/*
779 	 * If the branch acquired continuation since we've looked at it -
780 	 * fine, it should all survive and (new) top doesn't belong to us.
781 	 */
782 	if (!partial->key && *partial->p)
783 		/* Writer: end */
784 		goto no_top;
785 	for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
786 		;
787 	/*
788 	 * OK, we've found the last block that must survive. The rest of our
789 	 * branch should be detached before unlocking. However, if that rest
790 	 * of branch is all ours and does not grow immediately from the inode
791 	 * it's easier to cheat and just decrement partial->p.
792 	 */
793 	if (p == chain + k - 1 && p > chain) {
794 		p->p--;
795 	} else {
796 		*top = *p->p;
797 		/* Nope, don't do this in ext4.  Must leave the tree intact */
798 #if 0
799 		*p->p = 0;
800 #endif
801 	}
802 	/* Writer: end */
803 
804 	while (partial > p) {
805 		brelse(partial->bh);
806 		partial--;
807 	}
808 no_top:
809 	return partial;
810 }
811 
812 /*
813  * Zero a number of block pointers in either an inode or an indirect block.
814  * If we restart the transaction we must again get write access to the
815  * indirect block for further modification.
816  *
817  * We release `count' blocks on disk, but (last - first) may be greater
818  * than `count' because there can be holes in there.
819  *
820  * Return 0 on success, 1 on invalid block range
821  * and < 0 on fatal error.
822  */
823 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
824 			     struct buffer_head *bh,
825 			     ext4_fsblk_t block_to_free,
826 			     unsigned long count, __le32 *first,
827 			     __le32 *last)
828 {
829 	__le32 *p;
830 	int	flags = EXT4_FREE_BLOCKS_VALIDATED;
831 	int	err;
832 
833 	if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
834 	    ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
835 		flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
836 	else if (ext4_should_journal_data(inode))
837 		flags |= EXT4_FREE_BLOCKS_FORGET;
838 
839 	if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
840 				   count)) {
841 		EXT4_ERROR_INODE(inode, "attempt to clear invalid "
842 				 "blocks %llu len %lu",
843 				 (unsigned long long) block_to_free, count);
844 		return 1;
845 	}
846 
847 	if (try_to_extend_transaction(handle, inode)) {
848 		if (bh) {
849 			BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
850 			err = ext4_handle_dirty_metadata(handle, inode, bh);
851 			if (unlikely(err))
852 				goto out_err;
853 		}
854 		err = ext4_mark_inode_dirty(handle, inode);
855 		if (unlikely(err))
856 			goto out_err;
857 		err = ext4_truncate_restart_trans(handle, inode,
858 					ext4_blocks_for_truncate(inode));
859 		if (unlikely(err))
860 			goto out_err;
861 		if (bh) {
862 			BUFFER_TRACE(bh, "retaking write access");
863 			err = ext4_journal_get_write_access(handle, bh);
864 			if (unlikely(err))
865 				goto out_err;
866 		}
867 	}
868 
869 	for (p = first; p < last; p++)
870 		*p = 0;
871 
872 	ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
873 	return 0;
874 out_err:
875 	ext4_std_error(inode->i_sb, err);
876 	return err;
877 }
878 
879 /**
880  * ext4_free_data - free a list of data blocks
881  * @handle:	handle for this transaction
882  * @inode:	inode we are dealing with
883  * @this_bh:	indirect buffer_head which contains *@first and *@last
884  * @first:	array of block numbers
885  * @last:	points immediately past the end of array
886  *
887  * We are freeing all blocks referred from that array (numbers are stored as
888  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
889  *
890  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
891  * blocks are contiguous then releasing them at one time will only affect one
892  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
893  * actually use a lot of journal space.
894  *
895  * @this_bh will be %NULL if @first and @last point into the inode's direct
896  * block pointers.
897  */
898 static void ext4_free_data(handle_t *handle, struct inode *inode,
899 			   struct buffer_head *this_bh,
900 			   __le32 *first, __le32 *last)
901 {
902 	ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
903 	unsigned long count = 0;	    /* Number of blocks in the run */
904 	__le32 *block_to_free_p = NULL;	    /* Pointer into inode/ind
905 					       corresponding to
906 					       block_to_free */
907 	ext4_fsblk_t nr;		    /* Current block # */
908 	__le32 *p;			    /* Pointer into inode/ind
909 					       for current block */
910 	int err = 0;
911 
912 	if (this_bh) {				/* For indirect block */
913 		BUFFER_TRACE(this_bh, "get_write_access");
914 		err = ext4_journal_get_write_access(handle, this_bh);
915 		/* Important: if we can't update the indirect pointers
916 		 * to the blocks, we can't free them. */
917 		if (err)
918 			return;
919 	}
920 
921 	for (p = first; p < last; p++) {
922 		nr = le32_to_cpu(*p);
923 		if (nr) {
924 			/* accumulate blocks to free if they're contiguous */
925 			if (count == 0) {
926 				block_to_free = nr;
927 				block_to_free_p = p;
928 				count = 1;
929 			} else if (nr == block_to_free + count) {
930 				count++;
931 			} else {
932 				err = ext4_clear_blocks(handle, inode, this_bh,
933 						        block_to_free, count,
934 						        block_to_free_p, p);
935 				if (err)
936 					break;
937 				block_to_free = nr;
938 				block_to_free_p = p;
939 				count = 1;
940 			}
941 		}
942 	}
943 
944 	if (!err && count > 0)
945 		err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
946 					count, block_to_free_p, p);
947 	if (err < 0)
948 		/* fatal error */
949 		return;
950 
951 	if (this_bh) {
952 		BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
953 
954 		/*
955 		 * The buffer head should have an attached journal head at this
956 		 * point. However, if the data is corrupted and an indirect
957 		 * block pointed to itself, it would have been detached when
958 		 * the block was cleared. Check for this instead of OOPSing.
959 		 */
960 		if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
961 			ext4_handle_dirty_metadata(handle, inode, this_bh);
962 		else
963 			EXT4_ERROR_INODE(inode,
964 					 "circular indirect block detected at "
965 					 "block %llu",
966 				(unsigned long long) this_bh->b_blocknr);
967 	}
968 }
969 
970 /**
971  *	ext4_free_branches - free an array of branches
972  *	@handle: JBD handle for this transaction
973  *	@inode:	inode we are dealing with
974  *	@parent_bh: the buffer_head which contains *@first and *@last
975  *	@first:	array of block numbers
976  *	@last:	pointer immediately past the end of array
977  *	@depth:	depth of the branches to free
978  *
979  *	We are freeing all blocks referred from these branches (numbers are
980  *	stored as little-endian 32-bit) and updating @inode->i_blocks
981  *	appropriately.
982  */
983 static void ext4_free_branches(handle_t *handle, struct inode *inode,
984 			       struct buffer_head *parent_bh,
985 			       __le32 *first, __le32 *last, int depth)
986 {
987 	ext4_fsblk_t nr;
988 	__le32 *p;
989 
990 	if (ext4_handle_is_aborted(handle))
991 		return;
992 
993 	if (depth--) {
994 		struct buffer_head *bh;
995 		int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
996 		p = last;
997 		while (--p >= first) {
998 			nr = le32_to_cpu(*p);
999 			if (!nr)
1000 				continue;		/* A hole */
1001 
1002 			if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1003 						   nr, 1)) {
1004 				EXT4_ERROR_INODE(inode,
1005 						 "invalid indirect mapped "
1006 						 "block %lu (level %d)",
1007 						 (unsigned long) nr, depth);
1008 				break;
1009 			}
1010 
1011 			/* Go read the buffer for the next level down */
1012 			bh = sb_bread(inode->i_sb, nr);
1013 
1014 			/*
1015 			 * A read failure? Report error and clear slot
1016 			 * (should be rare).
1017 			 */
1018 			if (!bh) {
1019 				EXT4_ERROR_INODE_BLOCK(inode, nr,
1020 						       "Read failure");
1021 				continue;
1022 			}
1023 
1024 			/* This zaps the entire block.  Bottom up. */
1025 			BUFFER_TRACE(bh, "free child branches");
1026 			ext4_free_branches(handle, inode, bh,
1027 					(__le32 *) bh->b_data,
1028 					(__le32 *) bh->b_data + addr_per_block,
1029 					depth);
1030 			brelse(bh);
1031 
1032 			/*
1033 			 * Everything below this this pointer has been
1034 			 * released.  Now let this top-of-subtree go.
1035 			 *
1036 			 * We want the freeing of this indirect block to be
1037 			 * atomic in the journal with the updating of the
1038 			 * bitmap block which owns it.  So make some room in
1039 			 * the journal.
1040 			 *
1041 			 * We zero the parent pointer *after* freeing its
1042 			 * pointee in the bitmaps, so if extend_transaction()
1043 			 * for some reason fails to put the bitmap changes and
1044 			 * the release into the same transaction, recovery
1045 			 * will merely complain about releasing a free block,
1046 			 * rather than leaking blocks.
1047 			 */
1048 			if (ext4_handle_is_aborted(handle))
1049 				return;
1050 			if (try_to_extend_transaction(handle, inode)) {
1051 				ext4_mark_inode_dirty(handle, inode);
1052 				ext4_truncate_restart_trans(handle, inode,
1053 					    ext4_blocks_for_truncate(inode));
1054 			}
1055 
1056 			/*
1057 			 * The forget flag here is critical because if
1058 			 * we are journaling (and not doing data
1059 			 * journaling), we have to make sure a revoke
1060 			 * record is written to prevent the journal
1061 			 * replay from overwriting the (former)
1062 			 * indirect block if it gets reallocated as a
1063 			 * data block.  This must happen in the same
1064 			 * transaction where the data blocks are
1065 			 * actually freed.
1066 			 */
1067 			ext4_free_blocks(handle, inode, NULL, nr, 1,
1068 					 EXT4_FREE_BLOCKS_METADATA|
1069 					 EXT4_FREE_BLOCKS_FORGET);
1070 
1071 			if (parent_bh) {
1072 				/*
1073 				 * The block which we have just freed is
1074 				 * pointed to by an indirect block: journal it
1075 				 */
1076 				BUFFER_TRACE(parent_bh, "get_write_access");
1077 				if (!ext4_journal_get_write_access(handle,
1078 								   parent_bh)){
1079 					*p = 0;
1080 					BUFFER_TRACE(parent_bh,
1081 					"call ext4_handle_dirty_metadata");
1082 					ext4_handle_dirty_metadata(handle,
1083 								   inode,
1084 								   parent_bh);
1085 				}
1086 			}
1087 		}
1088 	} else {
1089 		/* We have reached the bottom of the tree. */
1090 		BUFFER_TRACE(parent_bh, "free data blocks");
1091 		ext4_free_data(handle, inode, parent_bh, first, last);
1092 	}
1093 }
1094 
1095 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1096 {
1097 	struct ext4_inode_info *ei = EXT4_I(inode);
1098 	__le32 *i_data = ei->i_data;
1099 	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1100 	ext4_lblk_t offsets[4];
1101 	Indirect chain[4];
1102 	Indirect *partial;
1103 	__le32 nr = 0;
1104 	int n = 0;
1105 	ext4_lblk_t last_block, max_block;
1106 	unsigned blocksize = inode->i_sb->s_blocksize;
1107 
1108 	last_block = (inode->i_size + blocksize-1)
1109 					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1110 	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1111 					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1112 
1113 	if (last_block != max_block) {
1114 		n = ext4_block_to_path(inode, last_block, offsets, NULL);
1115 		if (n == 0)
1116 			return;
1117 	}
1118 
1119 	ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1120 
1121 	/*
1122 	 * The orphan list entry will now protect us from any crash which
1123 	 * occurs before the truncate completes, so it is now safe to propagate
1124 	 * the new, shorter inode size (held for now in i_size) into the
1125 	 * on-disk inode. We do this via i_disksize, which is the value which
1126 	 * ext4 *really* writes onto the disk inode.
1127 	 */
1128 	ei->i_disksize = inode->i_size;
1129 
1130 	if (last_block == max_block) {
1131 		/*
1132 		 * It is unnecessary to free any data blocks if last_block is
1133 		 * equal to the indirect block limit.
1134 		 */
1135 		return;
1136 	} else if (n == 1) {		/* direct blocks */
1137 		ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1138 			       i_data + EXT4_NDIR_BLOCKS);
1139 		goto do_indirects;
1140 	}
1141 
1142 	partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1143 	/* Kill the top of shared branch (not detached) */
1144 	if (nr) {
1145 		if (partial == chain) {
1146 			/* Shared branch grows from the inode */
1147 			ext4_free_branches(handle, inode, NULL,
1148 					   &nr, &nr+1, (chain+n-1) - partial);
1149 			*partial->p = 0;
1150 			/*
1151 			 * We mark the inode dirty prior to restart,
1152 			 * and prior to stop.  No need for it here.
1153 			 */
1154 		} else {
1155 			/* Shared branch grows from an indirect block */
1156 			BUFFER_TRACE(partial->bh, "get_write_access");
1157 			ext4_free_branches(handle, inode, partial->bh,
1158 					partial->p,
1159 					partial->p+1, (chain+n-1) - partial);
1160 		}
1161 	}
1162 	/* Clear the ends of indirect blocks on the shared branch */
1163 	while (partial > chain) {
1164 		ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1165 				   (__le32*)partial->bh->b_data+addr_per_block,
1166 				   (chain+n-1) - partial);
1167 		BUFFER_TRACE(partial->bh, "call brelse");
1168 		brelse(partial->bh);
1169 		partial--;
1170 	}
1171 do_indirects:
1172 	/* Kill the remaining (whole) subtrees */
1173 	switch (offsets[0]) {
1174 	default:
1175 		nr = i_data[EXT4_IND_BLOCK];
1176 		if (nr) {
1177 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1178 			i_data[EXT4_IND_BLOCK] = 0;
1179 		}
1180 		/* fall through */
1181 	case EXT4_IND_BLOCK:
1182 		nr = i_data[EXT4_DIND_BLOCK];
1183 		if (nr) {
1184 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1185 			i_data[EXT4_DIND_BLOCK] = 0;
1186 		}
1187 		/* fall through */
1188 	case EXT4_DIND_BLOCK:
1189 		nr = i_data[EXT4_TIND_BLOCK];
1190 		if (nr) {
1191 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1192 			i_data[EXT4_TIND_BLOCK] = 0;
1193 		}
1194 		/* fall through */
1195 	case EXT4_TIND_BLOCK:
1196 		;
1197 	}
1198 }
1199 
1200 /**
1201  *	ext4_ind_remove_space - remove space from the range
1202  *	@handle: JBD handle for this transaction
1203  *	@inode:	inode we are dealing with
1204  *	@start:	First block to remove
1205  *	@end:	One block after the last block to remove (exclusive)
1206  *
1207  *	Free the blocks in the defined range (end is exclusive endpoint of
1208  *	range). This is used by ext4_punch_hole().
1209  */
1210 int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1211 			  ext4_lblk_t start, ext4_lblk_t end)
1212 {
1213 	struct ext4_inode_info *ei = EXT4_I(inode);
1214 	__le32 *i_data = ei->i_data;
1215 	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1216 	ext4_lblk_t offsets[4], offsets2[4];
1217 	Indirect chain[4], chain2[4];
1218 	Indirect *partial, *partial2;
1219 	Indirect *p = NULL, *p2 = NULL;
1220 	ext4_lblk_t max_block;
1221 	__le32 nr = 0, nr2 = 0;
1222 	int n = 0, n2 = 0;
1223 	unsigned blocksize = inode->i_sb->s_blocksize;
1224 
1225 	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1226 					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1227 	if (end >= max_block)
1228 		end = max_block;
1229 	if ((start >= end) || (start > max_block))
1230 		return 0;
1231 
1232 	n = ext4_block_to_path(inode, start, offsets, NULL);
1233 	n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1234 
1235 	BUG_ON(n > n2);
1236 
1237 	if ((n == 1) && (n == n2)) {
1238 		/* We're punching only within direct block range */
1239 		ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1240 			       i_data + offsets2[0]);
1241 		return 0;
1242 	} else if (n2 > n) {
1243 		/*
1244 		 * Start and end are on a different levels so we're going to
1245 		 * free partial block at start, and partial block at end of
1246 		 * the range. If there are some levels in between then
1247 		 * do_indirects label will take care of that.
1248 		 */
1249 
1250 		if (n == 1) {
1251 			/*
1252 			 * Start is at the direct block level, free
1253 			 * everything to the end of the level.
1254 			 */
1255 			ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1256 				       i_data + EXT4_NDIR_BLOCKS);
1257 			goto end_range;
1258 		}
1259 
1260 
1261 		partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1262 		if (nr) {
1263 			if (partial == chain) {
1264 				/* Shared branch grows from the inode */
1265 				ext4_free_branches(handle, inode, NULL,
1266 					   &nr, &nr+1, (chain+n-1) - partial);
1267 				*partial->p = 0;
1268 			} else {
1269 				/* Shared branch grows from an indirect block */
1270 				BUFFER_TRACE(partial->bh, "get_write_access");
1271 				ext4_free_branches(handle, inode, partial->bh,
1272 					partial->p,
1273 					partial->p+1, (chain+n-1) - partial);
1274 			}
1275 		}
1276 
1277 		/*
1278 		 * Clear the ends of indirect blocks on the shared branch
1279 		 * at the start of the range
1280 		 */
1281 		while (partial > chain) {
1282 			ext4_free_branches(handle, inode, partial->bh,
1283 				partial->p + 1,
1284 				(__le32 *)partial->bh->b_data+addr_per_block,
1285 				(chain+n-1) - partial);
1286 			partial--;
1287 		}
1288 
1289 end_range:
1290 		partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1291 		if (nr2) {
1292 			if (partial2 == chain2) {
1293 				/*
1294 				 * Remember, end is exclusive so here we're at
1295 				 * the start of the next level we're not going
1296 				 * to free. Everything was covered by the start
1297 				 * of the range.
1298 				 */
1299 				goto do_indirects;
1300 			}
1301 		} else {
1302 			/*
1303 			 * ext4_find_shared returns Indirect structure which
1304 			 * points to the last element which should not be
1305 			 * removed by truncate. But this is end of the range
1306 			 * in punch_hole so we need to point to the next element
1307 			 */
1308 			partial2->p++;
1309 		}
1310 
1311 		/*
1312 		 * Clear the ends of indirect blocks on the shared branch
1313 		 * at the end of the range
1314 		 */
1315 		while (partial2 > chain2) {
1316 			ext4_free_branches(handle, inode, partial2->bh,
1317 					   (__le32 *)partial2->bh->b_data,
1318 					   partial2->p,
1319 					   (chain2+n2-1) - partial2);
1320 			partial2--;
1321 		}
1322 		goto do_indirects;
1323 	}
1324 
1325 	/* Punch happened within the same level (n == n2) */
1326 	partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1327 	partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1328 
1329 	/* Free top, but only if partial2 isn't its subtree. */
1330 	if (nr) {
1331 		int level = min(partial - chain, partial2 - chain2);
1332 		int i;
1333 		int subtree = 1;
1334 
1335 		for (i = 0; i <= level; i++) {
1336 			if (offsets[i] != offsets2[i]) {
1337 				subtree = 0;
1338 				break;
1339 			}
1340 		}
1341 
1342 		if (!subtree) {
1343 			if (partial == chain) {
1344 				/* Shared branch grows from the inode */
1345 				ext4_free_branches(handle, inode, NULL,
1346 						   &nr, &nr+1,
1347 						   (chain+n-1) - partial);
1348 				*partial->p = 0;
1349 			} else {
1350 				/* Shared branch grows from an indirect block */
1351 				BUFFER_TRACE(partial->bh, "get_write_access");
1352 				ext4_free_branches(handle, inode, partial->bh,
1353 						   partial->p,
1354 						   partial->p+1,
1355 						   (chain+n-1) - partial);
1356 			}
1357 		}
1358 	}
1359 
1360 	if (!nr2) {
1361 		/*
1362 		 * ext4_find_shared returns Indirect structure which
1363 		 * points to the last element which should not be
1364 		 * removed by truncate. But this is end of the range
1365 		 * in punch_hole so we need to point to the next element
1366 		 */
1367 		partial2->p++;
1368 	}
1369 
1370 	while (partial > chain || partial2 > chain2) {
1371 		int depth = (chain+n-1) - partial;
1372 		int depth2 = (chain2+n2-1) - partial2;
1373 
1374 		if (partial > chain && partial2 > chain2 &&
1375 		    partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1376 			/*
1377 			 * We've converged on the same block. Clear the range,
1378 			 * then we're done.
1379 			 */
1380 			ext4_free_branches(handle, inode, partial->bh,
1381 					   partial->p + 1,
1382 					   partial2->p,
1383 					   (chain+n-1) - partial);
1384 			goto cleanup;
1385 		}
1386 
1387 		/*
1388 		 * The start and end partial branches may not be at the same
1389 		 * level even though the punch happened within one level. So, we
1390 		 * give them a chance to arrive at the same level, then walk
1391 		 * them in step with each other until we converge on the same
1392 		 * block.
1393 		 */
1394 		if (partial > chain && depth <= depth2) {
1395 			ext4_free_branches(handle, inode, partial->bh,
1396 					   partial->p + 1,
1397 					   (__le32 *)partial->bh->b_data+addr_per_block,
1398 					   (chain+n-1) - partial);
1399 			partial--;
1400 		}
1401 		if (partial2 > chain2 && depth2 <= depth) {
1402 			ext4_free_branches(handle, inode, partial2->bh,
1403 					   (__le32 *)partial2->bh->b_data,
1404 					   partial2->p,
1405 					   (chain2+n2-1) - partial2);
1406 			partial2--;
1407 		}
1408 	}
1409 
1410 cleanup:
1411 	while (p && p > chain) {
1412 		BUFFER_TRACE(p->bh, "call brelse");
1413 		brelse(p->bh);
1414 		p--;
1415 	}
1416 	while (p2 && p2 > chain2) {
1417 		BUFFER_TRACE(p2->bh, "call brelse");
1418 		brelse(p2->bh);
1419 		p2--;
1420 	}
1421 	return 0;
1422 
1423 do_indirects:
1424 	/* Kill the remaining (whole) subtrees */
1425 	switch (offsets[0]) {
1426 	default:
1427 		if (++n >= n2)
1428 			break;
1429 		nr = i_data[EXT4_IND_BLOCK];
1430 		if (nr) {
1431 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1432 			i_data[EXT4_IND_BLOCK] = 0;
1433 		}
1434 		/* fall through */
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 		/* fall through */
1444 	case EXT4_DIND_BLOCK:
1445 		if (++n >= n2)
1446 			break;
1447 		nr = i_data[EXT4_TIND_BLOCK];
1448 		if (nr) {
1449 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1450 			i_data[EXT4_TIND_BLOCK] = 0;
1451 		}
1452 		/* fall through */
1453 	case EXT4_TIND_BLOCK:
1454 		;
1455 	}
1456 	goto cleanup;
1457 }
1458