xref: /linux/fs/ext4/indirect.c (revision 148f9bb87745ed45f7a11b2cbd3bc0f017d5d257)
1 /*
2  *  linux/fs/ext4/indirect.c
3  *
4  *  from
5  *
6  *  linux/fs/ext4/inode.c
7  *
8  * Copyright (C) 1992, 1993, 1994, 1995
9  * Remy Card (card@masi.ibp.fr)
10  * Laboratoire MASI - Institut Blaise Pascal
11  * Universite Pierre et Marie Curie (Paris VI)
12  *
13  *  from
14  *
15  *  linux/fs/minix/inode.c
16  *
17  *  Copyright (C) 1991, 1992  Linus Torvalds
18  *
19  *  Goal-directed block allocation by Stephen Tweedie
20  *	(sct@redhat.com), 1993, 1998
21  */
22 
23 #include <linux/aio.h>
24 #include "ext4_jbd2.h"
25 #include "truncate.h"
26 #include "ext4_extents.h"	/* Needed for EXT_MAX_BLOCKS */
27 
28 #include <trace/events/ext4.h>
29 
30 typedef struct {
31 	__le32	*p;
32 	__le32	key;
33 	struct buffer_head *bh;
34 } Indirect;
35 
36 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
37 {
38 	p->key = *(p->p = v);
39 	p->bh = bh;
40 }
41 
42 /**
43  *	ext4_block_to_path - parse the block number into array of offsets
44  *	@inode: inode in question (we are only interested in its superblock)
45  *	@i_block: block number to be parsed
46  *	@offsets: array to store the offsets in
47  *	@boundary: set this non-zero if the referred-to block is likely to be
48  *	       followed (on disk) by an indirect block.
49  *
50  *	To store the locations of file's data ext4 uses a data structure common
51  *	for UNIX filesystems - tree of pointers anchored in the inode, with
52  *	data blocks at leaves and indirect blocks in intermediate nodes.
53  *	This function translates the block number into path in that tree -
54  *	return value is the path length and @offsets[n] is the offset of
55  *	pointer to (n+1)th node in the nth one. If @block is out of range
56  *	(negative or too large) warning is printed and zero returned.
57  *
58  *	Note: function doesn't find node addresses, so no IO is needed. All
59  *	we need to know is the capacity of indirect blocks (taken from the
60  *	inode->i_sb).
61  */
62 
63 /*
64  * Portability note: the last comparison (check that we fit into triple
65  * indirect block) is spelled differently, because otherwise on an
66  * architecture with 32-bit longs and 8Kb pages we might get into trouble
67  * if our filesystem had 8Kb blocks. We might use long long, but that would
68  * kill us on x86. Oh, well, at least the sign propagation does not matter -
69  * i_block would have to be negative in the very beginning, so we would not
70  * get there at all.
71  */
72 
73 static int ext4_block_to_path(struct inode *inode,
74 			      ext4_lblk_t i_block,
75 			      ext4_lblk_t offsets[4], int *boundary)
76 {
77 	int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
78 	int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
79 	const long direct_blocks = EXT4_NDIR_BLOCKS,
80 		indirect_blocks = ptrs,
81 		double_blocks = (1 << (ptrs_bits * 2));
82 	int n = 0;
83 	int final = 0;
84 
85 	if (i_block < direct_blocks) {
86 		offsets[n++] = i_block;
87 		final = direct_blocks;
88 	} else if ((i_block -= direct_blocks) < indirect_blocks) {
89 		offsets[n++] = EXT4_IND_BLOCK;
90 		offsets[n++] = i_block;
91 		final = ptrs;
92 	} else if ((i_block -= indirect_blocks) < double_blocks) {
93 		offsets[n++] = EXT4_DIND_BLOCK;
94 		offsets[n++] = i_block >> ptrs_bits;
95 		offsets[n++] = i_block & (ptrs - 1);
96 		final = ptrs;
97 	} else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
98 		offsets[n++] = EXT4_TIND_BLOCK;
99 		offsets[n++] = i_block >> (ptrs_bits * 2);
100 		offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
101 		offsets[n++] = i_block & (ptrs - 1);
102 		final = ptrs;
103 	} else {
104 		ext4_warning(inode->i_sb, "block %lu > max in inode %lu",
105 			     i_block + direct_blocks +
106 			     indirect_blocks + double_blocks, inode->i_ino);
107 	}
108 	if (boundary)
109 		*boundary = final - 1 - (i_block & (ptrs - 1));
110 	return n;
111 }
112 
113 /**
114  *	ext4_get_branch - read the chain of indirect blocks leading to data
115  *	@inode: inode in question
116  *	@depth: depth of the chain (1 - direct pointer, etc.)
117  *	@offsets: offsets of pointers in inode/indirect blocks
118  *	@chain: place to store the result
119  *	@err: here we store the error value
120  *
121  *	Function fills the array of triples <key, p, bh> and returns %NULL
122  *	if everything went OK or the pointer to the last filled triple
123  *	(incomplete one) otherwise. Upon the return chain[i].key contains
124  *	the number of (i+1)-th block in the chain (as it is stored in memory,
125  *	i.e. little-endian 32-bit), chain[i].p contains the address of that
126  *	number (it points into struct inode for i==0 and into the bh->b_data
127  *	for i>0) and chain[i].bh points to the buffer_head of i-th indirect
128  *	block for i>0 and NULL for i==0. In other words, it holds the block
129  *	numbers of the chain, addresses they were taken from (and where we can
130  *	verify that chain did not change) and buffer_heads hosting these
131  *	numbers.
132  *
133  *	Function stops when it stumbles upon zero pointer (absent block)
134  *		(pointer to last triple returned, *@err == 0)
135  *	or when it gets an IO error reading an indirect block
136  *		(ditto, *@err == -EIO)
137  *	or when it reads all @depth-1 indirect blocks successfully and finds
138  *	the whole chain, all way to the data (returns %NULL, *err == 0).
139  *
140  *      Need to be called with
141  *      down_read(&EXT4_I(inode)->i_data_sem)
142  */
143 static Indirect *ext4_get_branch(struct inode *inode, int depth,
144 				 ext4_lblk_t  *offsets,
145 				 Indirect chain[4], int *err)
146 {
147 	struct super_block *sb = inode->i_sb;
148 	Indirect *p = chain;
149 	struct buffer_head *bh;
150 	int ret = -EIO;
151 
152 	*err = 0;
153 	/* i_data is not going away, no lock needed */
154 	add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
155 	if (!p->key)
156 		goto no_block;
157 	while (--depth) {
158 		bh = sb_getblk(sb, le32_to_cpu(p->key));
159 		if (unlikely(!bh)) {
160 			ret = -ENOMEM;
161 			goto failure;
162 		}
163 
164 		if (!bh_uptodate_or_lock(bh)) {
165 			if (bh_submit_read(bh) < 0) {
166 				put_bh(bh);
167 				goto failure;
168 			}
169 			/* validate block references */
170 			if (ext4_check_indirect_blockref(inode, bh)) {
171 				put_bh(bh);
172 				goto failure;
173 			}
174 		}
175 
176 		add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
177 		/* Reader: end */
178 		if (!p->key)
179 			goto no_block;
180 	}
181 	return NULL;
182 
183 failure:
184 	*err = ret;
185 no_block:
186 	return p;
187 }
188 
189 /**
190  *	ext4_find_near - find a place for allocation with sufficient locality
191  *	@inode: owner
192  *	@ind: descriptor of indirect block.
193  *
194  *	This function returns the preferred place for block allocation.
195  *	It is used when heuristic for sequential allocation fails.
196  *	Rules are:
197  *	  + if there is a block to the left of our position - allocate near it.
198  *	  + if pointer will live in indirect block - allocate near that block.
199  *	  + if pointer will live in inode - allocate in the same
200  *	    cylinder group.
201  *
202  * In the latter case we colour the starting block by the callers PID to
203  * prevent it from clashing with concurrent allocations for a different inode
204  * in the same block group.   The PID is used here so that functionally related
205  * files will be close-by on-disk.
206  *
207  *	Caller must make sure that @ind is valid and will stay that way.
208  */
209 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
210 {
211 	struct ext4_inode_info *ei = EXT4_I(inode);
212 	__le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
213 	__le32 *p;
214 
215 	/* Try to find previous block */
216 	for (p = ind->p - 1; p >= start; p--) {
217 		if (*p)
218 			return le32_to_cpu(*p);
219 	}
220 
221 	/* No such thing, so let's try location of indirect block */
222 	if (ind->bh)
223 		return ind->bh->b_blocknr;
224 
225 	/*
226 	 * It is going to be referred to from the inode itself? OK, just put it
227 	 * into the same cylinder group then.
228 	 */
229 	return ext4_inode_to_goal_block(inode);
230 }
231 
232 /**
233  *	ext4_find_goal - find a preferred place for allocation.
234  *	@inode: owner
235  *	@block:  block we want
236  *	@partial: pointer to the last triple within a chain
237  *
238  *	Normally this function find the preferred place for block allocation,
239  *	returns it.
240  *	Because this is only used for non-extent files, we limit the block nr
241  *	to 32 bits.
242  */
243 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
244 				   Indirect *partial)
245 {
246 	ext4_fsblk_t goal;
247 
248 	/*
249 	 * XXX need to get goal block from mballoc's data structures
250 	 */
251 
252 	goal = ext4_find_near(inode, partial);
253 	goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
254 	return goal;
255 }
256 
257 /**
258  *	ext4_blks_to_allocate - Look up the block map and count the number
259  *	of direct blocks need to be allocated for the given branch.
260  *
261  *	@branch: chain of indirect blocks
262  *	@k: number of blocks need for indirect blocks
263  *	@blks: number of data blocks to be mapped.
264  *	@blocks_to_boundary:  the offset in the indirect block
265  *
266  *	return the total number of blocks to be allocate, including the
267  *	direct and indirect blocks.
268  */
269 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
270 				 int blocks_to_boundary)
271 {
272 	unsigned int count = 0;
273 
274 	/*
275 	 * Simple case, [t,d]Indirect block(s) has not allocated yet
276 	 * then it's clear blocks on that path have not allocated
277 	 */
278 	if (k > 0) {
279 		/* right now we don't handle cross boundary allocation */
280 		if (blks < blocks_to_boundary + 1)
281 			count += blks;
282 		else
283 			count += blocks_to_boundary + 1;
284 		return count;
285 	}
286 
287 	count++;
288 	while (count < blks && count <= blocks_to_boundary &&
289 		le32_to_cpu(*(branch[0].p + count)) == 0) {
290 		count++;
291 	}
292 	return count;
293 }
294 
295 /**
296  *	ext4_alloc_branch - allocate and set up a chain of blocks.
297  *	@handle: handle for this transaction
298  *	@inode: owner
299  *	@indirect_blks: number of allocated indirect blocks
300  *	@blks: number of allocated direct blocks
301  *	@goal: preferred place for allocation
302  *	@offsets: offsets (in the blocks) to store the pointers to next.
303  *	@branch: place to store the chain in.
304  *
305  *	This function allocates blocks, zeroes out all but the last one,
306  *	links them into chain and (if we are synchronous) writes them to disk.
307  *	In other words, it prepares a branch that can be spliced onto the
308  *	inode. It stores the information about that chain in the branch[], in
309  *	the same format as ext4_get_branch() would do. We are calling it after
310  *	we had read the existing part of chain and partial points to the last
311  *	triple of that (one with zero ->key). Upon the exit we have the same
312  *	picture as after the successful ext4_get_block(), except that in one
313  *	place chain is disconnected - *branch->p is still zero (we did not
314  *	set the last link), but branch->key contains the number that should
315  *	be placed into *branch->p to fill that gap.
316  *
317  *	If allocation fails we free all blocks we've allocated (and forget
318  *	their buffer_heads) and return the error value the from failed
319  *	ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
320  *	as described above and return 0.
321  */
322 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
323 			     ext4_lblk_t iblock, int indirect_blks,
324 			     int *blks, ext4_fsblk_t goal,
325 			     ext4_lblk_t *offsets, Indirect *branch)
326 {
327 	struct ext4_allocation_request	ar;
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 	/*
334 	 * Set up for the direct block allocation
335 	 */
336 	memset(&ar, 0, sizeof(ar));
337 	ar.inode = inode;
338 	ar.len = *blks;
339 	ar.logical = iblock;
340 	if (S_ISREG(inode->i_mode))
341 		ar.flags = EXT4_MB_HINT_DATA;
342 
343 	for (i = 0; i <= indirect_blks; i++) {
344 		if (i == indirect_blks) {
345 			ar.goal = goal;
346 			new_blocks[i] = ext4_mb_new_blocks(handle, &ar, &err);
347 		} else
348 			goal = new_blocks[i] = ext4_new_meta_blocks(handle, inode,
349 							goal, 0, NULL, &err);
350 		if (err) {
351 			i--;
352 			goto failed;
353 		}
354 		branch[i].key = cpu_to_le32(new_blocks[i]);
355 		if (i == 0)
356 			continue;
357 
358 		bh = branch[i].bh = sb_getblk(inode->i_sb, new_blocks[i-1]);
359 		if (unlikely(!bh)) {
360 			err = -ENOMEM;
361 			goto failed;
362 		}
363 		lock_buffer(bh);
364 		BUFFER_TRACE(bh, "call get_create_access");
365 		err = ext4_journal_get_create_access(handle, bh);
366 		if (err) {
367 			unlock_buffer(bh);
368 			goto failed;
369 		}
370 
371 		memset(bh->b_data, 0, bh->b_size);
372 		p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
373 		b = new_blocks[i];
374 
375 		if (i == indirect_blks)
376 			len = ar.len;
377 		for (j = 0; j < len; j++)
378 			*p++ = cpu_to_le32(b++);
379 
380 		BUFFER_TRACE(bh, "marking uptodate");
381 		set_buffer_uptodate(bh);
382 		unlock_buffer(bh);
383 
384 		BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
385 		err = ext4_handle_dirty_metadata(handle, inode, bh);
386 		if (err)
387 			goto failed;
388 	}
389 	*blks = ar.len;
390 	return 0;
391 failed:
392 	for (; i >= 0; i--) {
393 		if (i != indirect_blks && branch[i].bh)
394 			ext4_forget(handle, 1, inode, branch[i].bh,
395 				    branch[i].bh->b_blocknr);
396 		ext4_free_blocks(handle, inode, NULL, new_blocks[i],
397 				 (i == indirect_blks) ? ar.len : 1, 0);
398 	}
399 	return err;
400 }
401 
402 /**
403  * ext4_splice_branch - splice the allocated branch onto inode.
404  * @handle: handle for this transaction
405  * @inode: owner
406  * @block: (logical) number of block we are adding
407  * @chain: chain of indirect blocks (with a missing link - see
408  *	ext4_alloc_branch)
409  * @where: location of missing link
410  * @num:   number of indirect blocks we are adding
411  * @blks:  number of direct blocks we are adding
412  *
413  * This function fills the missing link and does all housekeeping needed in
414  * inode (->i_blocks, etc.). In case of success we end up with the full
415  * chain to new block and return 0.
416  */
417 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
418 			      ext4_lblk_t block, Indirect *where, int num,
419 			      int blks)
420 {
421 	int i;
422 	int err = 0;
423 	ext4_fsblk_t current_block;
424 
425 	/*
426 	 * If we're splicing into a [td]indirect block (as opposed to the
427 	 * inode) then we need to get write access to the [td]indirect block
428 	 * before the splice.
429 	 */
430 	if (where->bh) {
431 		BUFFER_TRACE(where->bh, "get_write_access");
432 		err = ext4_journal_get_write_access(handle, where->bh);
433 		if (err)
434 			goto err_out;
435 	}
436 	/* That's it */
437 
438 	*where->p = where->key;
439 
440 	/*
441 	 * Update the host buffer_head or inode to point to more just allocated
442 	 * direct blocks blocks
443 	 */
444 	if (num == 0 && blks > 1) {
445 		current_block = le32_to_cpu(where->key) + 1;
446 		for (i = 1; i < blks; i++)
447 			*(where->p + i) = cpu_to_le32(current_block++);
448 	}
449 
450 	/* We are done with atomic stuff, now do the rest of housekeeping */
451 	/* had we spliced it onto indirect block? */
452 	if (where->bh) {
453 		/*
454 		 * If we spliced it onto an indirect block, we haven't
455 		 * altered the inode.  Note however that if it is being spliced
456 		 * onto an indirect block at the very end of the file (the
457 		 * file is growing) then we *will* alter the inode to reflect
458 		 * the new i_size.  But that is not done here - it is done in
459 		 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
460 		 */
461 		jbd_debug(5, "splicing indirect only\n");
462 		BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
463 		err = ext4_handle_dirty_metadata(handle, inode, where->bh);
464 		if (err)
465 			goto err_out;
466 	} else {
467 		/*
468 		 * OK, we spliced it into the inode itself on a direct block.
469 		 */
470 		ext4_mark_inode_dirty(handle, inode);
471 		jbd_debug(5, "splicing direct\n");
472 	}
473 	return err;
474 
475 err_out:
476 	for (i = 1; i <= num; i++) {
477 		/*
478 		 * branch[i].bh is newly allocated, so there is no
479 		 * need to revoke the block, which is why we don't
480 		 * need to set EXT4_FREE_BLOCKS_METADATA.
481 		 */
482 		ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
483 				 EXT4_FREE_BLOCKS_FORGET);
484 	}
485 	ext4_free_blocks(handle, inode, NULL, le32_to_cpu(where[num].key),
486 			 blks, 0);
487 
488 	return err;
489 }
490 
491 /*
492  * The ext4_ind_map_blocks() function handles non-extents inodes
493  * (i.e., using the traditional indirect/double-indirect i_blocks
494  * scheme) for ext4_map_blocks().
495  *
496  * Allocation strategy is simple: if we have to allocate something, we will
497  * have to go the whole way to leaf. So let's do it before attaching anything
498  * to tree, set linkage between the newborn blocks, write them if sync is
499  * required, recheck the path, free and repeat if check fails, otherwise
500  * set the last missing link (that will protect us from any truncate-generated
501  * removals - all blocks on the path are immune now) and possibly force the
502  * write on the parent block.
503  * That has a nice additional property: no special recovery from the failed
504  * allocations is needed - we simply release blocks and do not touch anything
505  * reachable from inode.
506  *
507  * `handle' can be NULL if create == 0.
508  *
509  * return > 0, # of blocks mapped or allocated.
510  * return = 0, if plain lookup failed.
511  * return < 0, error case.
512  *
513  * The ext4_ind_get_blocks() function should be called with
514  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
515  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
516  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
517  * blocks.
518  */
519 int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
520 			struct ext4_map_blocks *map,
521 			int flags)
522 {
523 	int err = -EIO;
524 	ext4_lblk_t offsets[4];
525 	Indirect chain[4];
526 	Indirect *partial;
527 	ext4_fsblk_t goal;
528 	int indirect_blks;
529 	int blocks_to_boundary = 0;
530 	int depth;
531 	int count = 0;
532 	ext4_fsblk_t first_block = 0;
533 
534 	trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
535 	J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
536 	J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
537 	depth = ext4_block_to_path(inode, map->m_lblk, offsets,
538 				   &blocks_to_boundary);
539 
540 	if (depth == 0)
541 		goto out;
542 
543 	partial = ext4_get_branch(inode, depth, offsets, chain, &err);
544 
545 	/* Simplest case - block found, no allocation needed */
546 	if (!partial) {
547 		first_block = le32_to_cpu(chain[depth - 1].key);
548 		count++;
549 		/*map more blocks*/
550 		while (count < map->m_len && count <= blocks_to_boundary) {
551 			ext4_fsblk_t blk;
552 
553 			blk = le32_to_cpu(*(chain[depth-1].p + count));
554 
555 			if (blk == first_block + count)
556 				count++;
557 			else
558 				break;
559 		}
560 		goto got_it;
561 	}
562 
563 	/* Next simple case - plain lookup or failed read of indirect block */
564 	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
565 		goto cleanup;
566 
567 	/*
568 	 * Okay, we need to do block allocation.
569 	*/
570 	if (EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
571 				       EXT4_FEATURE_RO_COMPAT_BIGALLOC)) {
572 		EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
573 				 "non-extent mapped inodes with bigalloc");
574 		return -ENOSPC;
575 	}
576 
577 	goal = ext4_find_goal(inode, map->m_lblk, partial);
578 
579 	/* the number of blocks need to allocate for [d,t]indirect blocks */
580 	indirect_blks = (chain + depth) - partial - 1;
581 
582 	/*
583 	 * Next look up the indirect map to count the totoal number of
584 	 * direct blocks to allocate for this branch.
585 	 */
586 	count = ext4_blks_to_allocate(partial, indirect_blks,
587 				      map->m_len, blocks_to_boundary);
588 	/*
589 	 * Block out ext4_truncate while we alter the tree
590 	 */
591 	err = ext4_alloc_branch(handle, inode, map->m_lblk, indirect_blks,
592 				&count, goal,
593 				offsets + (partial - chain), partial);
594 
595 	/*
596 	 * The ext4_splice_branch call will free and forget any buffers
597 	 * on the new chain if there is a failure, but that risks using
598 	 * up transaction credits, especially for bitmaps where the
599 	 * credits cannot be returned.  Can we handle this somehow?  We
600 	 * may need to return -EAGAIN upwards in the worst case.  --sct
601 	 */
602 	if (!err)
603 		err = ext4_splice_branch(handle, inode, map->m_lblk,
604 					 partial, indirect_blks, count);
605 	if (err)
606 		goto cleanup;
607 
608 	map->m_flags |= EXT4_MAP_NEW;
609 
610 	ext4_update_inode_fsync_trans(handle, inode, 1);
611 got_it:
612 	map->m_flags |= EXT4_MAP_MAPPED;
613 	map->m_pblk = le32_to_cpu(chain[depth-1].key);
614 	map->m_len = count;
615 	if (count > blocks_to_boundary)
616 		map->m_flags |= EXT4_MAP_BOUNDARY;
617 	err = count;
618 	/* Clean up and exit */
619 	partial = chain + depth - 1;	/* the whole chain */
620 cleanup:
621 	while (partial > chain) {
622 		BUFFER_TRACE(partial->bh, "call brelse");
623 		brelse(partial->bh);
624 		partial--;
625 	}
626 out:
627 	trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
628 	return err;
629 }
630 
631 /*
632  * O_DIRECT for ext3 (or indirect map) based files
633  *
634  * If the O_DIRECT write will extend the file then add this inode to the
635  * orphan list.  So recovery will truncate it back to the original size
636  * if the machine crashes during the write.
637  *
638  * If the O_DIRECT write is intantiating holes inside i_size and the machine
639  * crashes then stale disk data _may_ be exposed inside the file. But current
640  * VFS code falls back into buffered path in that case so we are safe.
641  */
642 ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
643 			   const struct iovec *iov, loff_t offset,
644 			   unsigned long nr_segs)
645 {
646 	struct file *file = iocb->ki_filp;
647 	struct inode *inode = file->f_mapping->host;
648 	struct ext4_inode_info *ei = EXT4_I(inode);
649 	handle_t *handle;
650 	ssize_t ret;
651 	int orphan = 0;
652 	size_t count = iov_length(iov, nr_segs);
653 	int retries = 0;
654 
655 	if (rw == WRITE) {
656 		loff_t final_size = offset + count;
657 
658 		if (final_size > inode->i_size) {
659 			/* Credits for sb + inode write */
660 			handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
661 			if (IS_ERR(handle)) {
662 				ret = PTR_ERR(handle);
663 				goto out;
664 			}
665 			ret = ext4_orphan_add(handle, inode);
666 			if (ret) {
667 				ext4_journal_stop(handle);
668 				goto out;
669 			}
670 			orphan = 1;
671 			ei->i_disksize = inode->i_size;
672 			ext4_journal_stop(handle);
673 		}
674 	}
675 
676 retry:
677 	if (rw == READ && ext4_should_dioread_nolock(inode)) {
678 		/*
679 		 * Nolock dioread optimization may be dynamically disabled
680 		 * via ext4_inode_block_unlocked_dio(). Check inode's state
681 		 * while holding extra i_dio_count ref.
682 		 */
683 		atomic_inc(&inode->i_dio_count);
684 		smp_mb();
685 		if (unlikely(ext4_test_inode_state(inode,
686 						    EXT4_STATE_DIOREAD_LOCK))) {
687 			inode_dio_done(inode);
688 			goto locked;
689 		}
690 		ret = __blockdev_direct_IO(rw, iocb, inode,
691 				 inode->i_sb->s_bdev, iov,
692 				 offset, nr_segs,
693 				 ext4_get_block, NULL, NULL, 0);
694 		inode_dio_done(inode);
695 	} else {
696 locked:
697 		ret = blockdev_direct_IO(rw, iocb, inode, iov,
698 				 offset, nr_segs, ext4_get_block);
699 
700 		if (unlikely((rw & WRITE) && ret < 0)) {
701 			loff_t isize = i_size_read(inode);
702 			loff_t end = offset + iov_length(iov, nr_segs);
703 
704 			if (end > isize)
705 				ext4_truncate_failed_write(inode);
706 		}
707 	}
708 	if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
709 		goto retry;
710 
711 	if (orphan) {
712 		int err;
713 
714 		/* Credits for sb + inode write */
715 		handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
716 		if (IS_ERR(handle)) {
717 			/* This is really bad luck. We've written the data
718 			 * but cannot extend i_size. Bail out and pretend
719 			 * the write failed... */
720 			ret = PTR_ERR(handle);
721 			if (inode->i_nlink)
722 				ext4_orphan_del(NULL, inode);
723 
724 			goto out;
725 		}
726 		if (inode->i_nlink)
727 			ext4_orphan_del(handle, inode);
728 		if (ret > 0) {
729 			loff_t end = offset + ret;
730 			if (end > inode->i_size) {
731 				ei->i_disksize = end;
732 				i_size_write(inode, end);
733 				/*
734 				 * We're going to return a positive `ret'
735 				 * here due to non-zero-length I/O, so there's
736 				 * no way of reporting error returns from
737 				 * ext4_mark_inode_dirty() to userspace.  So
738 				 * ignore it.
739 				 */
740 				ext4_mark_inode_dirty(handle, inode);
741 			}
742 		}
743 		err = ext4_journal_stop(handle);
744 		if (ret == 0)
745 			ret = err;
746 	}
747 out:
748 	return ret;
749 }
750 
751 /*
752  * Calculate the number of metadata blocks need to reserve
753  * to allocate a new block at @lblocks for non extent file based file
754  */
755 int ext4_ind_calc_metadata_amount(struct inode *inode, sector_t lblock)
756 {
757 	struct ext4_inode_info *ei = EXT4_I(inode);
758 	sector_t dind_mask = ~((sector_t)EXT4_ADDR_PER_BLOCK(inode->i_sb) - 1);
759 	int blk_bits;
760 
761 	if (lblock < EXT4_NDIR_BLOCKS)
762 		return 0;
763 
764 	lblock -= EXT4_NDIR_BLOCKS;
765 
766 	if (ei->i_da_metadata_calc_len &&
767 	    (lblock & dind_mask) == ei->i_da_metadata_calc_last_lblock) {
768 		ei->i_da_metadata_calc_len++;
769 		return 0;
770 	}
771 	ei->i_da_metadata_calc_last_lblock = lblock & dind_mask;
772 	ei->i_da_metadata_calc_len = 1;
773 	blk_bits = order_base_2(lblock);
774 	return (blk_bits / EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb)) + 1;
775 }
776 
777 /*
778  * Calculate number of indirect blocks touched by mapping @nrblocks logically
779  * contiguous blocks
780  */
781 int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
782 {
783 	/*
784 	 * With N contiguous data blocks, we need at most
785 	 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
786 	 * 2 dindirect blocks, and 1 tindirect block
787 	 */
788 	return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
789 }
790 
791 /*
792  * Truncate transactions can be complex and absolutely huge.  So we need to
793  * be able to restart the transaction at a conventient checkpoint to make
794  * sure we don't overflow the journal.
795  *
796  * Try to extend this transaction for the purposes of truncation.  If
797  * extend fails, we need to propagate the failure up and restart the
798  * transaction in the top-level truncate loop. --sct
799  *
800  * Returns 0 if we managed to create more room.  If we can't create more
801  * room, and the transaction must be restarted we return 1.
802  */
803 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
804 {
805 	if (!ext4_handle_valid(handle))
806 		return 0;
807 	if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
808 		return 0;
809 	if (!ext4_journal_extend(handle, ext4_blocks_for_truncate(inode)))
810 		return 0;
811 	return 1;
812 }
813 
814 /*
815  * Probably it should be a library function... search for first non-zero word
816  * or memcmp with zero_page, whatever is better for particular architecture.
817  * Linus?
818  */
819 static inline int all_zeroes(__le32 *p, __le32 *q)
820 {
821 	while (p < q)
822 		if (*p++)
823 			return 0;
824 	return 1;
825 }
826 
827 /**
828  *	ext4_find_shared - find the indirect blocks for partial truncation.
829  *	@inode:	  inode in question
830  *	@depth:	  depth of the affected branch
831  *	@offsets: offsets of pointers in that branch (see ext4_block_to_path)
832  *	@chain:	  place to store the pointers to partial indirect blocks
833  *	@top:	  place to the (detached) top of branch
834  *
835  *	This is a helper function used by ext4_truncate().
836  *
837  *	When we do truncate() we may have to clean the ends of several
838  *	indirect blocks but leave the blocks themselves alive. Block is
839  *	partially truncated if some data below the new i_size is referred
840  *	from it (and it is on the path to the first completely truncated
841  *	data block, indeed).  We have to free the top of that path along
842  *	with everything to the right of the path. Since no allocation
843  *	past the truncation point is possible until ext4_truncate()
844  *	finishes, we may safely do the latter, but top of branch may
845  *	require special attention - pageout below the truncation point
846  *	might try to populate it.
847  *
848  *	We atomically detach the top of branch from the tree, store the
849  *	block number of its root in *@top, pointers to buffer_heads of
850  *	partially truncated blocks - in @chain[].bh and pointers to
851  *	their last elements that should not be removed - in
852  *	@chain[].p. Return value is the pointer to last filled element
853  *	of @chain.
854  *
855  *	The work left to caller to do the actual freeing of subtrees:
856  *		a) free the subtree starting from *@top
857  *		b) free the subtrees whose roots are stored in
858  *			(@chain[i].p+1 .. end of @chain[i].bh->b_data)
859  *		c) free the subtrees growing from the inode past the @chain[0].
860  *			(no partially truncated stuff there).  */
861 
862 static Indirect *ext4_find_shared(struct inode *inode, int depth,
863 				  ext4_lblk_t offsets[4], Indirect chain[4],
864 				  __le32 *top)
865 {
866 	Indirect *partial, *p;
867 	int k, err;
868 
869 	*top = 0;
870 	/* Make k index the deepest non-null offset + 1 */
871 	for (k = depth; k > 1 && !offsets[k-1]; k--)
872 		;
873 	partial = ext4_get_branch(inode, k, offsets, chain, &err);
874 	/* Writer: pointers */
875 	if (!partial)
876 		partial = chain + k-1;
877 	/*
878 	 * If the branch acquired continuation since we've looked at it -
879 	 * fine, it should all survive and (new) top doesn't belong to us.
880 	 */
881 	if (!partial->key && *partial->p)
882 		/* Writer: end */
883 		goto no_top;
884 	for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
885 		;
886 	/*
887 	 * OK, we've found the last block that must survive. The rest of our
888 	 * branch should be detached before unlocking. However, if that rest
889 	 * of branch is all ours and does not grow immediately from the inode
890 	 * it's easier to cheat and just decrement partial->p.
891 	 */
892 	if (p == chain + k - 1 && p > chain) {
893 		p->p--;
894 	} else {
895 		*top = *p->p;
896 		/* Nope, don't do this in ext4.  Must leave the tree intact */
897 #if 0
898 		*p->p = 0;
899 #endif
900 	}
901 	/* Writer: end */
902 
903 	while (partial > p) {
904 		brelse(partial->bh);
905 		partial--;
906 	}
907 no_top:
908 	return partial;
909 }
910 
911 /*
912  * Zero a number of block pointers in either an inode or an indirect block.
913  * If we restart the transaction we must again get write access to the
914  * indirect block for further modification.
915  *
916  * We release `count' blocks on disk, but (last - first) may be greater
917  * than `count' because there can be holes in there.
918  *
919  * Return 0 on success, 1 on invalid block range
920  * and < 0 on fatal error.
921  */
922 static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
923 			     struct buffer_head *bh,
924 			     ext4_fsblk_t block_to_free,
925 			     unsigned long count, __le32 *first,
926 			     __le32 *last)
927 {
928 	__le32 *p;
929 	int	flags = EXT4_FREE_BLOCKS_VALIDATED;
930 	int	err;
931 
932 	if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
933 		flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
934 	else if (ext4_should_journal_data(inode))
935 		flags |= EXT4_FREE_BLOCKS_FORGET;
936 
937 	if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), block_to_free,
938 				   count)) {
939 		EXT4_ERROR_INODE(inode, "attempt to clear invalid "
940 				 "blocks %llu len %lu",
941 				 (unsigned long long) block_to_free, count);
942 		return 1;
943 	}
944 
945 	if (try_to_extend_transaction(handle, inode)) {
946 		if (bh) {
947 			BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
948 			err = ext4_handle_dirty_metadata(handle, inode, bh);
949 			if (unlikely(err))
950 				goto out_err;
951 		}
952 		err = ext4_mark_inode_dirty(handle, inode);
953 		if (unlikely(err))
954 			goto out_err;
955 		err = ext4_truncate_restart_trans(handle, inode,
956 					ext4_blocks_for_truncate(inode));
957 		if (unlikely(err))
958 			goto out_err;
959 		if (bh) {
960 			BUFFER_TRACE(bh, "retaking write access");
961 			err = ext4_journal_get_write_access(handle, bh);
962 			if (unlikely(err))
963 				goto out_err;
964 		}
965 	}
966 
967 	for (p = first; p < last; p++)
968 		*p = 0;
969 
970 	ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
971 	return 0;
972 out_err:
973 	ext4_std_error(inode->i_sb, err);
974 	return err;
975 }
976 
977 /**
978  * ext4_free_data - free a list of data blocks
979  * @handle:	handle for this transaction
980  * @inode:	inode we are dealing with
981  * @this_bh:	indirect buffer_head which contains *@first and *@last
982  * @first:	array of block numbers
983  * @last:	points immediately past the end of array
984  *
985  * We are freeing all blocks referred from that array (numbers are stored as
986  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
987  *
988  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
989  * blocks are contiguous then releasing them at one time will only affect one
990  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
991  * actually use a lot of journal space.
992  *
993  * @this_bh will be %NULL if @first and @last point into the inode's direct
994  * block pointers.
995  */
996 static void ext4_free_data(handle_t *handle, struct inode *inode,
997 			   struct buffer_head *this_bh,
998 			   __le32 *first, __le32 *last)
999 {
1000 	ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
1001 	unsigned long count = 0;	    /* Number of blocks in the run */
1002 	__le32 *block_to_free_p = NULL;	    /* Pointer into inode/ind
1003 					       corresponding to
1004 					       block_to_free */
1005 	ext4_fsblk_t nr;		    /* Current block # */
1006 	__le32 *p;			    /* Pointer into inode/ind
1007 					       for current block */
1008 	int err = 0;
1009 
1010 	if (this_bh) {				/* For indirect block */
1011 		BUFFER_TRACE(this_bh, "get_write_access");
1012 		err = ext4_journal_get_write_access(handle, this_bh);
1013 		/* Important: if we can't update the indirect pointers
1014 		 * to the blocks, we can't free them. */
1015 		if (err)
1016 			return;
1017 	}
1018 
1019 	for (p = first; p < last; p++) {
1020 		nr = le32_to_cpu(*p);
1021 		if (nr) {
1022 			/* accumulate blocks to free if they're contiguous */
1023 			if (count == 0) {
1024 				block_to_free = nr;
1025 				block_to_free_p = p;
1026 				count = 1;
1027 			} else if (nr == block_to_free + count) {
1028 				count++;
1029 			} else {
1030 				err = ext4_clear_blocks(handle, inode, this_bh,
1031 						        block_to_free, count,
1032 						        block_to_free_p, p);
1033 				if (err)
1034 					break;
1035 				block_to_free = nr;
1036 				block_to_free_p = p;
1037 				count = 1;
1038 			}
1039 		}
1040 	}
1041 
1042 	if (!err && count > 0)
1043 		err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
1044 					count, block_to_free_p, p);
1045 	if (err < 0)
1046 		/* fatal error */
1047 		return;
1048 
1049 	if (this_bh) {
1050 		BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
1051 
1052 		/*
1053 		 * The buffer head should have an attached journal head at this
1054 		 * point. However, if the data is corrupted and an indirect
1055 		 * block pointed to itself, it would have been detached when
1056 		 * the block was cleared. Check for this instead of OOPSing.
1057 		 */
1058 		if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
1059 			ext4_handle_dirty_metadata(handle, inode, this_bh);
1060 		else
1061 			EXT4_ERROR_INODE(inode,
1062 					 "circular indirect block detected at "
1063 					 "block %llu",
1064 				(unsigned long long) this_bh->b_blocknr);
1065 	}
1066 }
1067 
1068 /**
1069  *	ext4_free_branches - free an array of branches
1070  *	@handle: JBD handle for this transaction
1071  *	@inode:	inode we are dealing with
1072  *	@parent_bh: the buffer_head which contains *@first and *@last
1073  *	@first:	array of block numbers
1074  *	@last:	pointer immediately past the end of array
1075  *	@depth:	depth of the branches to free
1076  *
1077  *	We are freeing all blocks referred from these branches (numbers are
1078  *	stored as little-endian 32-bit) and updating @inode->i_blocks
1079  *	appropriately.
1080  */
1081 static void ext4_free_branches(handle_t *handle, struct inode *inode,
1082 			       struct buffer_head *parent_bh,
1083 			       __le32 *first, __le32 *last, int depth)
1084 {
1085 	ext4_fsblk_t nr;
1086 	__le32 *p;
1087 
1088 	if (ext4_handle_is_aborted(handle))
1089 		return;
1090 
1091 	if (depth--) {
1092 		struct buffer_head *bh;
1093 		int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1094 		p = last;
1095 		while (--p >= first) {
1096 			nr = le32_to_cpu(*p);
1097 			if (!nr)
1098 				continue;		/* A hole */
1099 
1100 			if (!ext4_data_block_valid(EXT4_SB(inode->i_sb),
1101 						   nr, 1)) {
1102 				EXT4_ERROR_INODE(inode,
1103 						 "invalid indirect mapped "
1104 						 "block %lu (level %d)",
1105 						 (unsigned long) nr, depth);
1106 				break;
1107 			}
1108 
1109 			/* Go read the buffer for the next level down */
1110 			bh = sb_bread(inode->i_sb, nr);
1111 
1112 			/*
1113 			 * A read failure? Report error and clear slot
1114 			 * (should be rare).
1115 			 */
1116 			if (!bh) {
1117 				EXT4_ERROR_INODE_BLOCK(inode, nr,
1118 						       "Read failure");
1119 				continue;
1120 			}
1121 
1122 			/* This zaps the entire block.  Bottom up. */
1123 			BUFFER_TRACE(bh, "free child branches");
1124 			ext4_free_branches(handle, inode, bh,
1125 					(__le32 *) bh->b_data,
1126 					(__le32 *) bh->b_data + addr_per_block,
1127 					depth);
1128 			brelse(bh);
1129 
1130 			/*
1131 			 * Everything below this this pointer has been
1132 			 * released.  Now let this top-of-subtree go.
1133 			 *
1134 			 * We want the freeing of this indirect block to be
1135 			 * atomic in the journal with the updating of the
1136 			 * bitmap block which owns it.  So make some room in
1137 			 * the journal.
1138 			 *
1139 			 * We zero the parent pointer *after* freeing its
1140 			 * pointee in the bitmaps, so if extend_transaction()
1141 			 * for some reason fails to put the bitmap changes and
1142 			 * the release into the same transaction, recovery
1143 			 * will merely complain about releasing a free block,
1144 			 * rather than leaking blocks.
1145 			 */
1146 			if (ext4_handle_is_aborted(handle))
1147 				return;
1148 			if (try_to_extend_transaction(handle, inode)) {
1149 				ext4_mark_inode_dirty(handle, inode);
1150 				ext4_truncate_restart_trans(handle, inode,
1151 					    ext4_blocks_for_truncate(inode));
1152 			}
1153 
1154 			/*
1155 			 * The forget flag here is critical because if
1156 			 * we are journaling (and not doing data
1157 			 * journaling), we have to make sure a revoke
1158 			 * record is written to prevent the journal
1159 			 * replay from overwriting the (former)
1160 			 * indirect block if it gets reallocated as a
1161 			 * data block.  This must happen in the same
1162 			 * transaction where the data blocks are
1163 			 * actually freed.
1164 			 */
1165 			ext4_free_blocks(handle, inode, NULL, nr, 1,
1166 					 EXT4_FREE_BLOCKS_METADATA|
1167 					 EXT4_FREE_BLOCKS_FORGET);
1168 
1169 			if (parent_bh) {
1170 				/*
1171 				 * The block which we have just freed is
1172 				 * pointed to by an indirect block: journal it
1173 				 */
1174 				BUFFER_TRACE(parent_bh, "get_write_access");
1175 				if (!ext4_journal_get_write_access(handle,
1176 								   parent_bh)){
1177 					*p = 0;
1178 					BUFFER_TRACE(parent_bh,
1179 					"call ext4_handle_dirty_metadata");
1180 					ext4_handle_dirty_metadata(handle,
1181 								   inode,
1182 								   parent_bh);
1183 				}
1184 			}
1185 		}
1186 	} else {
1187 		/* We have reached the bottom of the tree. */
1188 		BUFFER_TRACE(parent_bh, "free data blocks");
1189 		ext4_free_data(handle, inode, parent_bh, first, last);
1190 	}
1191 }
1192 
1193 void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1194 {
1195 	struct ext4_inode_info *ei = EXT4_I(inode);
1196 	__le32 *i_data = ei->i_data;
1197 	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1198 	ext4_lblk_t offsets[4];
1199 	Indirect chain[4];
1200 	Indirect *partial;
1201 	__le32 nr = 0;
1202 	int n = 0;
1203 	ext4_lblk_t last_block, max_block;
1204 	unsigned blocksize = inode->i_sb->s_blocksize;
1205 
1206 	last_block = (inode->i_size + blocksize-1)
1207 					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1208 	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1209 					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1210 
1211 	if (last_block != max_block) {
1212 		n = ext4_block_to_path(inode, last_block, offsets, NULL);
1213 		if (n == 0)
1214 			return;
1215 	}
1216 
1217 	ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1218 
1219 	/*
1220 	 * The orphan list entry will now protect us from any crash which
1221 	 * occurs before the truncate completes, so it is now safe to propagate
1222 	 * the new, shorter inode size (held for now in i_size) into the
1223 	 * on-disk inode. We do this via i_disksize, which is the value which
1224 	 * ext4 *really* writes onto the disk inode.
1225 	 */
1226 	ei->i_disksize = inode->i_size;
1227 
1228 	if (last_block == max_block) {
1229 		/*
1230 		 * It is unnecessary to free any data blocks if last_block is
1231 		 * equal to the indirect block limit.
1232 		 */
1233 		return;
1234 	} else if (n == 1) {		/* direct blocks */
1235 		ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1236 			       i_data + EXT4_NDIR_BLOCKS);
1237 		goto do_indirects;
1238 	}
1239 
1240 	partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1241 	/* Kill the top of shared branch (not detached) */
1242 	if (nr) {
1243 		if (partial == chain) {
1244 			/* Shared branch grows from the inode */
1245 			ext4_free_branches(handle, inode, NULL,
1246 					   &nr, &nr+1, (chain+n-1) - partial);
1247 			*partial->p = 0;
1248 			/*
1249 			 * We mark the inode dirty prior to restart,
1250 			 * and prior to stop.  No need for it here.
1251 			 */
1252 		} else {
1253 			/* Shared branch grows from an indirect block */
1254 			BUFFER_TRACE(partial->bh, "get_write_access");
1255 			ext4_free_branches(handle, inode, partial->bh,
1256 					partial->p,
1257 					partial->p+1, (chain+n-1) - partial);
1258 		}
1259 	}
1260 	/* Clear the ends of indirect blocks on the shared branch */
1261 	while (partial > chain) {
1262 		ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1263 				   (__le32*)partial->bh->b_data+addr_per_block,
1264 				   (chain+n-1) - partial);
1265 		BUFFER_TRACE(partial->bh, "call brelse");
1266 		brelse(partial->bh);
1267 		partial--;
1268 	}
1269 do_indirects:
1270 	/* Kill the remaining (whole) subtrees */
1271 	switch (offsets[0]) {
1272 	default:
1273 		nr = i_data[EXT4_IND_BLOCK];
1274 		if (nr) {
1275 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1276 			i_data[EXT4_IND_BLOCK] = 0;
1277 		}
1278 	case EXT4_IND_BLOCK:
1279 		nr = i_data[EXT4_DIND_BLOCK];
1280 		if (nr) {
1281 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1282 			i_data[EXT4_DIND_BLOCK] = 0;
1283 		}
1284 	case EXT4_DIND_BLOCK:
1285 		nr = i_data[EXT4_TIND_BLOCK];
1286 		if (nr) {
1287 			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1288 			i_data[EXT4_TIND_BLOCK] = 0;
1289 		}
1290 	case EXT4_TIND_BLOCK:
1291 		;
1292 	}
1293 }
1294 
1295 static int free_hole_blocks(handle_t *handle, struct inode *inode,
1296 			    struct buffer_head *parent_bh, __le32 *i_data,
1297 			    int level, ext4_lblk_t first,
1298 			    ext4_lblk_t count, int max)
1299 {
1300 	struct buffer_head *bh = NULL;
1301 	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1302 	int ret = 0;
1303 	int i, inc;
1304 	ext4_lblk_t offset;
1305 	__le32 blk;
1306 
1307 	inc = 1 << ((EXT4_BLOCK_SIZE_BITS(inode->i_sb) - 2) * level);
1308 	for (i = 0, offset = 0; i < max; i++, i_data++, offset += inc) {
1309 		if (offset >= count + first)
1310 			break;
1311 		if (*i_data == 0 || (offset + inc) <= first)
1312 			continue;
1313 		blk = *i_data;
1314 		if (level > 0) {
1315 			ext4_lblk_t first2;
1316 			bh = sb_bread(inode->i_sb, le32_to_cpu(blk));
1317 			if (!bh) {
1318 				EXT4_ERROR_INODE_BLOCK(inode, le32_to_cpu(blk),
1319 						       "Read failure");
1320 				return -EIO;
1321 			}
1322 			first2 = (first > offset) ? first - offset : 0;
1323 			ret = free_hole_blocks(handle, inode, bh,
1324 					       (__le32 *)bh->b_data, level - 1,
1325 					       first2, count - offset,
1326 					       inode->i_sb->s_blocksize >> 2);
1327 			if (ret) {
1328 				brelse(bh);
1329 				goto err;
1330 			}
1331 		}
1332 		if (level == 0 ||
1333 		    (bh && all_zeroes((__le32 *)bh->b_data,
1334 				      (__le32 *)bh->b_data + addr_per_block))) {
1335 			ext4_free_data(handle, inode, parent_bh, &blk, &blk+1);
1336 			*i_data = 0;
1337 		}
1338 		brelse(bh);
1339 		bh = NULL;
1340 	}
1341 
1342 err:
1343 	return ret;
1344 }
1345 
1346 int ext4_free_hole_blocks(handle_t *handle, struct inode *inode,
1347 			  ext4_lblk_t first, ext4_lblk_t stop)
1348 {
1349 	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1350 	int level, ret = 0;
1351 	int num = EXT4_NDIR_BLOCKS;
1352 	ext4_lblk_t count, max = EXT4_NDIR_BLOCKS;
1353 	__le32 *i_data = EXT4_I(inode)->i_data;
1354 
1355 	count = stop - first;
1356 	for (level = 0; level < 4; level++, max *= addr_per_block) {
1357 		if (first < max) {
1358 			ret = free_hole_blocks(handle, inode, NULL, i_data,
1359 					       level, first, count, num);
1360 			if (ret)
1361 				goto err;
1362 			if (count > max - first)
1363 				count -= max - first;
1364 			else
1365 				break;
1366 			first = 0;
1367 		} else {
1368 			first -= max;
1369 		}
1370 		i_data += num;
1371 		if (level == 0) {
1372 			num = 1;
1373 			max = 1;
1374 		}
1375 	}
1376 
1377 err:
1378 	return ret;
1379 }
1380 
1381