xref: /kernel/linux/linux-5.10/fs/ext4/indirect.c (revision 8c2ecf20)
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
31typedef struct {
32	__le32	*p;
33	__le32	key;
34	struct buffer_head *bh;
35} Indirect;
36
37static 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
74static 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 */
144static Indirect *ext4_get_branch(struct inode *inode, int depth,
145				 ext4_lblk_t  *offsets,
146				 Indirect chain[4], int *err)
147{
148	struct super_block *sb = inode->i_sb;
149	Indirect *p = chain;
150	struct buffer_head *bh;
151	unsigned int key;
152	int ret = -EIO;
153
154	*err = 0;
155	/* i_data is not going away, no lock needed */
156	add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
157	if (!p->key)
158		goto no_block;
159	while (--depth) {
160		key = le32_to_cpu(p->key);
161		if (key > ext4_blocks_count(EXT4_SB(sb)->s_es)) {
162			/* the block was out of range */
163			ret = -EFSCORRUPTED;
164			goto failure;
165		}
166		bh = sb_getblk(sb, key);
167		if (unlikely(!bh)) {
168			ret = -ENOMEM;
169			goto failure;
170		}
171
172		if (!bh_uptodate_or_lock(bh)) {
173			if (ext4_read_bh(bh, 0, NULL) < 0) {
174				put_bh(bh);
175				goto failure;
176			}
177			/* validate block references */
178			if (ext4_check_indirect_blockref(inode, bh)) {
179				put_bh(bh);
180				goto failure;
181			}
182		}
183
184		add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
185		/* Reader: end */
186		if (!p->key)
187			goto no_block;
188	}
189	return NULL;
190
191failure:
192	*err = ret;
193no_block:
194	return p;
195}
196
197/**
198 *	ext4_find_near - find a place for allocation with sufficient locality
199 *	@inode: owner
200 *	@ind: descriptor of indirect block.
201 *
202 *	This function returns the preferred place for block allocation.
203 *	It is used when heuristic for sequential allocation fails.
204 *	Rules are:
205 *	  + if there is a block to the left of our position - allocate near it.
206 *	  + if pointer will live in indirect block - allocate near that block.
207 *	  + if pointer will live in inode - allocate in the same
208 *	    cylinder group.
209 *
210 * In the latter case we colour the starting block by the callers PID to
211 * prevent it from clashing with concurrent allocations for a different inode
212 * in the same block group.   The PID is used here so that functionally related
213 * files will be close-by on-disk.
214 *
215 *	Caller must make sure that @ind is valid and will stay that way.
216 */
217static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
218{
219	struct ext4_inode_info *ei = EXT4_I(inode);
220	__le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
221	__le32 *p;
222
223	/* Try to find previous block */
224	for (p = ind->p - 1; p >= start; p--) {
225		if (*p)
226			return le32_to_cpu(*p);
227	}
228
229	/* No such thing, so let's try location of indirect block */
230	if (ind->bh)
231		return ind->bh->b_blocknr;
232
233	/*
234	 * It is going to be referred to from the inode itself? OK, just put it
235	 * into the same cylinder group then.
236	 */
237	return ext4_inode_to_goal_block(inode);
238}
239
240/**
241 *	ext4_find_goal - find a preferred place for allocation.
242 *	@inode: owner
243 *	@block:  block we want
244 *	@partial: pointer to the last triple within a chain
245 *
246 *	Normally this function find the preferred place for block allocation,
247 *	returns it.
248 *	Because this is only used for non-extent files, we limit the block nr
249 *	to 32 bits.
250 */
251static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
252				   Indirect *partial)
253{
254	ext4_fsblk_t goal;
255
256	/*
257	 * XXX need to get goal block from mballoc's data structures
258	 */
259
260	goal = ext4_find_near(inode, partial);
261	goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
262	return goal;
263}
264
265/**
266 *	ext4_blks_to_allocate - Look up the block map and count the number
267 *	of direct blocks need to be allocated for the given branch.
268 *
269 *	@branch: chain of indirect blocks
270 *	@k: number of blocks need for indirect blocks
271 *	@blks: number of data blocks to be mapped.
272 *	@blocks_to_boundary:  the offset in the indirect block
273 *
274 *	return the total number of blocks to be allocate, including the
275 *	direct and indirect blocks.
276 */
277static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
278				 int blocks_to_boundary)
279{
280	unsigned int count = 0;
281
282	/*
283	 * Simple case, [t,d]Indirect block(s) has not allocated yet
284	 * then it's clear blocks on that path have not allocated
285	 */
286	if (k > 0) {
287		/* right now we don't handle cross boundary allocation */
288		if (blks < blocks_to_boundary + 1)
289			count += blks;
290		else
291			count += blocks_to_boundary + 1;
292		return count;
293	}
294
295	count++;
296	while (count < blks && count <= blocks_to_boundary &&
297		le32_to_cpu(*(branch[0].p + count)) == 0) {
298		count++;
299	}
300	return count;
301}
302
303/**
304 * ext4_alloc_branch() - allocate and set up a chain of blocks
305 * @handle: handle for this transaction
306 * @ar: structure describing the allocation request
307 * @indirect_blks: number of allocated indirect blocks
308 * @offsets: offsets (in the blocks) to store the pointers to next.
309 * @branch: place to store the chain in.
310 *
311 *	This function allocates blocks, zeroes out all but the last one,
312 *	links them into chain and (if we are synchronous) writes them to disk.
313 *	In other words, it prepares a branch that can be spliced onto the
314 *	inode. It stores the information about that chain in the branch[], in
315 *	the same format as ext4_get_branch() would do. We are calling it after
316 *	we had read the existing part of chain and partial points to the last
317 *	triple of that (one with zero ->key). Upon the exit we have the same
318 *	picture as after the successful ext4_get_block(), except that in one
319 *	place chain is disconnected - *branch->p is still zero (we did not
320 *	set the last link), but branch->key contains the number that should
321 *	be placed into *branch->p to fill that gap.
322 *
323 *	If allocation fails we free all blocks we've allocated (and forget
324 *	their buffer_heads) and return the error value the from failed
325 *	ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
326 *	as described above and return 0.
327 */
328static int ext4_alloc_branch(handle_t *handle,
329			     struct ext4_allocation_request *ar,
330			     int indirect_blks, ext4_lblk_t *offsets,
331			     Indirect *branch)
332{
333	struct buffer_head *		bh;
334	ext4_fsblk_t			b, new_blocks[4];
335	__le32				*p;
336	int				i, j, err, len = 1;
337
338	for (i = 0; i <= indirect_blks; i++) {
339		if (i == indirect_blks) {
340			new_blocks[i] = ext4_mb_new_blocks(handle, ar, &err);
341		} else {
342			ar->goal = new_blocks[i] = ext4_new_meta_blocks(handle,
343					ar->inode, ar->goal,
344					ar->flags & EXT4_MB_DELALLOC_RESERVED,
345					NULL, &err);
346			/* Simplify error cleanup... */
347			branch[i+1].bh = NULL;
348		}
349		if (err) {
350			i--;
351			goto failed;
352		}
353		branch[i].key = cpu_to_le32(new_blocks[i]);
354		if (i == 0)
355			continue;
356
357		bh = branch[i].bh = sb_getblk(ar->inode->i_sb, new_blocks[i-1]);
358		if (unlikely(!bh)) {
359			err = -ENOMEM;
360			goto failed;
361		}
362		lock_buffer(bh);
363		BUFFER_TRACE(bh, "call get_create_access");
364		err = ext4_journal_get_create_access(handle, bh);
365		if (err) {
366			unlock_buffer(bh);
367			goto failed;
368		}
369
370		memset(bh->b_data, 0, bh->b_size);
371		p = branch[i].p = (__le32 *) bh->b_data + offsets[i];
372		b = new_blocks[i];
373
374		if (i == indirect_blks)
375			len = ar->len;
376		for (j = 0; j < len; j++)
377			*p++ = cpu_to_le32(b++);
378
379		BUFFER_TRACE(bh, "marking uptodate");
380		set_buffer_uptodate(bh);
381		unlock_buffer(bh);
382
383		BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
384		err = ext4_handle_dirty_metadata(handle, ar->inode, bh);
385		if (err)
386			goto failed;
387	}
388	return 0;
389failed:
390	if (i == indirect_blks) {
391		/* Free data blocks */
392		ext4_free_blocks(handle, ar->inode, NULL, new_blocks[i],
393				 ar->len, 0);
394		i--;
395	}
396	for (; i >= 0; i--) {
397		/*
398		 * We want to ext4_forget() only freshly allocated indirect
399		 * blocks. Buffer for new_blocks[i] is at branch[i+1].bh
400		 * (buffer at branch[0].bh is indirect block / inode already
401		 * existing before ext4_alloc_branch() was called). Also
402		 * because blocks are freshly allocated, we don't need to
403		 * revoke them which is why we don't set
404		 * EXT4_FREE_BLOCKS_METADATA.
405		 */
406		ext4_free_blocks(handle, ar->inode, branch[i+1].bh,
407				 new_blocks[i], 1,
408				 branch[i+1].bh ? EXT4_FREE_BLOCKS_FORGET : 0);
409	}
410	return err;
411}
412
413/**
414 * ext4_splice_branch() - splice the allocated branch onto inode.
415 * @handle: handle for this transaction
416 * @ar: structure describing the allocation request
417 * @where: location of missing link
418 * @num:   number of indirect blocks we are adding
419 *
420 * This function fills the missing link and does all housekeeping needed in
421 * inode (->i_blocks, etc.). In case of success we end up with the full
422 * chain to new block and return 0.
423 */
424static int ext4_splice_branch(handle_t *handle,
425			      struct ext4_allocation_request *ar,
426			      Indirect *where, int num)
427{
428	int i;
429	int err = 0;
430	ext4_fsblk_t current_block;
431
432	/*
433	 * If we're splicing into a [td]indirect block (as opposed to the
434	 * inode) then we need to get write access to the [td]indirect block
435	 * before the splice.
436	 */
437	if (where->bh) {
438		BUFFER_TRACE(where->bh, "get_write_access");
439		err = ext4_journal_get_write_access(handle, where->bh);
440		if (err)
441			goto err_out;
442	}
443	/* That's it */
444
445	*where->p = where->key;
446
447	/*
448	 * Update the host buffer_head or inode to point to more just allocated
449	 * direct blocks blocks
450	 */
451	if (num == 0 && ar->len > 1) {
452		current_block = le32_to_cpu(where->key) + 1;
453		for (i = 1; i < ar->len; i++)
454			*(where->p + i) = cpu_to_le32(current_block++);
455	}
456
457	/* We are done with atomic stuff, now do the rest of housekeeping */
458	/* had we spliced it onto indirect block? */
459	if (where->bh) {
460		/*
461		 * If we spliced it onto an indirect block, we haven't
462		 * altered the inode.  Note however that if it is being spliced
463		 * onto an indirect block at the very end of the file (the
464		 * file is growing) then we *will* alter the inode to reflect
465		 * the new i_size.  But that is not done here - it is done in
466		 * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
467		 */
468		jbd_debug(5, "splicing indirect only\n");
469		BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
470		err = ext4_handle_dirty_metadata(handle, ar->inode, where->bh);
471		if (err)
472			goto err_out;
473	} else {
474		/*
475		 * OK, we spliced it into the inode itself on a direct block.
476		 */
477		err = ext4_mark_inode_dirty(handle, ar->inode);
478		if (unlikely(err))
479			goto err_out;
480		jbd_debug(5, "splicing direct\n");
481	}
482	return err;
483
484err_out:
485	for (i = 1; i <= num; i++) {
486		/*
487		 * branch[i].bh is newly allocated, so there is no
488		 * need to revoke the block, which is why we don't
489		 * need to set EXT4_FREE_BLOCKS_METADATA.
490		 */
491		ext4_free_blocks(handle, ar->inode, where[i].bh, 0, 1,
492				 EXT4_FREE_BLOCKS_FORGET);
493	}
494	ext4_free_blocks(handle, ar->inode, NULL, le32_to_cpu(where[num].key),
495			 ar->len, 0);
496
497	return err;
498}
499
500/*
501 * The ext4_ind_map_blocks() function handles non-extents inodes
502 * (i.e., using the traditional indirect/double-indirect i_blocks
503 * scheme) for ext4_map_blocks().
504 *
505 * Allocation strategy is simple: if we have to allocate something, we will
506 * have to go the whole way to leaf. So let's do it before attaching anything
507 * to tree, set linkage between the newborn blocks, write them if sync is
508 * required, recheck the path, free and repeat if check fails, otherwise
509 * set the last missing link (that will protect us from any truncate-generated
510 * removals - all blocks on the path are immune now) and possibly force the
511 * write on the parent block.
512 * That has a nice additional property: no special recovery from the failed
513 * allocations is needed - we simply release blocks and do not touch anything
514 * reachable from inode.
515 *
516 * `handle' can be NULL if create == 0.
517 *
518 * return > 0, # of blocks mapped or allocated.
519 * return = 0, if plain lookup failed.
520 * return < 0, error case.
521 *
522 * The ext4_ind_get_blocks() function should be called with
523 * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
524 * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
525 * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
526 * blocks.
527 */
528int ext4_ind_map_blocks(handle_t *handle, struct inode *inode,
529			struct ext4_map_blocks *map,
530			int flags)
531{
532	struct ext4_allocation_request ar;
533	int err = -EIO;
534	ext4_lblk_t offsets[4];
535	Indirect chain[4];
536	Indirect *partial;
537	int indirect_blks;
538	int blocks_to_boundary = 0;
539	int depth;
540	int count = 0;
541	ext4_fsblk_t first_block = 0;
542
543	trace_ext4_ind_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
544	J_ASSERT(!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)));
545	J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
546	depth = ext4_block_to_path(inode, map->m_lblk, offsets,
547				   &blocks_to_boundary);
548
549	if (depth == 0)
550		goto out;
551
552	partial = ext4_get_branch(inode, depth, offsets, chain, &err);
553
554	/* Simplest case - block found, no allocation needed */
555	if (!partial) {
556		first_block = le32_to_cpu(chain[depth - 1].key);
557		count++;
558		/*map more blocks*/
559		while (count < map->m_len && count <= blocks_to_boundary) {
560			ext4_fsblk_t blk;
561
562			blk = le32_to_cpu(*(chain[depth-1].p + count));
563
564			if (blk == first_block + count)
565				count++;
566			else
567				break;
568		}
569		goto got_it;
570	}
571
572	/* Next simple case - plain lookup failed */
573	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
574		unsigned epb = inode->i_sb->s_blocksize / sizeof(u32);
575		int i;
576
577		/*
578		 * Count number blocks in a subtree under 'partial'. At each
579		 * level we count number of complete empty subtrees beyond
580		 * current offset and then descend into the subtree only
581		 * partially beyond current offset.
582		 */
583		count = 0;
584		for (i = partial - chain + 1; i < depth; i++)
585			count = count * epb + (epb - offsets[i] - 1);
586		count++;
587		/* Fill in size of a hole we found */
588		map->m_pblk = 0;
589		map->m_len = min_t(unsigned int, map->m_len, count);
590		goto cleanup;
591	}
592
593	/* Failed read of indirect block */
594	if (err == -EIO)
595		goto cleanup;
596
597	/*
598	 * Okay, we need to do block allocation.
599	*/
600	if (ext4_has_feature_bigalloc(inode->i_sb)) {
601		EXT4_ERROR_INODE(inode, "Can't allocate blocks for "
602				 "non-extent mapped inodes with bigalloc");
603		err = -EFSCORRUPTED;
604		goto out;
605	}
606
607	/* Set up for the direct block allocation */
608	memset(&ar, 0, sizeof(ar));
609	ar.inode = inode;
610	ar.logical = map->m_lblk;
611	if (S_ISREG(inode->i_mode))
612		ar.flags = EXT4_MB_HINT_DATA;
613	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
614		ar.flags |= EXT4_MB_DELALLOC_RESERVED;
615	if (flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
616		ar.flags |= EXT4_MB_USE_RESERVED;
617
618	ar.goal = ext4_find_goal(inode, map->m_lblk, partial);
619
620	/* the number of blocks need to allocate for [d,t]indirect blocks */
621	indirect_blks = (chain + depth) - partial - 1;
622
623	/*
624	 * Next look up the indirect map to count the totoal number of
625	 * direct blocks to allocate for this branch.
626	 */
627	ar.len = ext4_blks_to_allocate(partial, indirect_blks,
628				       map->m_len, blocks_to_boundary);
629
630	/*
631	 * Block out ext4_truncate while we alter the tree
632	 */
633	err = ext4_alloc_branch(handle, &ar, indirect_blks,
634				offsets + (partial - chain), partial);
635
636	/*
637	 * The ext4_splice_branch call will free and forget any buffers
638	 * on the new chain if there is a failure, but that risks using
639	 * up transaction credits, especially for bitmaps where the
640	 * credits cannot be returned.  Can we handle this somehow?  We
641	 * may need to return -EAGAIN upwards in the worst case.  --sct
642	 */
643	if (!err)
644		err = ext4_splice_branch(handle, &ar, partial, indirect_blks);
645	if (err)
646		goto cleanup;
647
648	map->m_flags |= EXT4_MAP_NEW;
649
650	ext4_update_inode_fsync_trans(handle, inode, 1);
651	count = ar.len;
652
653	/*
654	 * Update reserved blocks/metadata blocks after successful block
655	 * allocation which had been deferred till now.
656	 */
657	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
658		ext4_da_update_reserve_space(inode, count, 1);
659
660got_it:
661	map->m_flags |= EXT4_MAP_MAPPED;
662	map->m_pblk = le32_to_cpu(chain[depth-1].key);
663	map->m_len = count;
664	if (count > blocks_to_boundary)
665		map->m_flags |= EXT4_MAP_BOUNDARY;
666	err = count;
667	/* Clean up and exit */
668	partial = chain + depth - 1;	/* the whole chain */
669cleanup:
670	while (partial > chain) {
671		BUFFER_TRACE(partial->bh, "call brelse");
672		brelse(partial->bh);
673		partial--;
674	}
675out:
676	trace_ext4_ind_map_blocks_exit(inode, flags, map, err);
677	return err;
678}
679
680/*
681 * Calculate number of indirect blocks touched by mapping @nrblocks logically
682 * contiguous blocks
683 */
684int ext4_ind_trans_blocks(struct inode *inode, int nrblocks)
685{
686	/*
687	 * With N contiguous data blocks, we need at most
688	 * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
689	 * 2 dindirect blocks, and 1 tindirect block
690	 */
691	return DIV_ROUND_UP(nrblocks, EXT4_ADDR_PER_BLOCK(inode->i_sb)) + 4;
692}
693
694static int ext4_ind_trunc_restart_fn(handle_t *handle, struct inode *inode,
695				     struct buffer_head *bh, int *dropped)
696{
697	int err;
698
699	if (bh) {
700		BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
701		err = ext4_handle_dirty_metadata(handle, inode, bh);
702		if (unlikely(err))
703			return err;
704	}
705	err = ext4_mark_inode_dirty(handle, inode);
706	if (unlikely(err))
707		return err;
708	/*
709	 * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
710	 * moment, get_block can be called only for blocks inside i_size since
711	 * page cache has been already dropped and writes are blocked by
712	 * i_mutex. So we can safely drop the i_data_sem here.
713	 */
714	BUG_ON(EXT4_JOURNAL(inode) == NULL);
715	ext4_discard_preallocations(inode, 0);
716	up_write(&EXT4_I(inode)->i_data_sem);
717	*dropped = 1;
718	return 0;
719}
720
721/*
722 * Truncate transactions can be complex and absolutely huge.  So we need to
723 * be able to restart the transaction at a convenient checkpoint to make
724 * sure we don't overflow the journal.
725 *
726 * Try to extend this transaction for the purposes of truncation.  If
727 * extend fails, we restart transaction.
728 */
729static int ext4_ind_truncate_ensure_credits(handle_t *handle,
730					    struct inode *inode,
731					    struct buffer_head *bh,
732					    int revoke_creds)
733{
734	int ret;
735	int dropped = 0;
736
737	ret = ext4_journal_ensure_credits_fn(handle, EXT4_RESERVE_TRANS_BLOCKS,
738			ext4_blocks_for_truncate(inode), revoke_creds,
739			ext4_ind_trunc_restart_fn(handle, inode, bh, &dropped));
740	if (dropped)
741		down_write(&EXT4_I(inode)->i_data_sem);
742	if (ret <= 0)
743		return ret;
744	if (bh) {
745		BUFFER_TRACE(bh, "retaking write access");
746		ret = ext4_journal_get_write_access(handle, bh);
747		if (unlikely(ret))
748			return ret;
749	}
750	return 0;
751}
752
753/*
754 * Probably it should be a library function... search for first non-zero word
755 * or memcmp with zero_page, whatever is better for particular architecture.
756 * Linus?
757 */
758static inline int all_zeroes(__le32 *p, __le32 *q)
759{
760	while (p < q)
761		if (*p++)
762			return 0;
763	return 1;
764}
765
766/**
767 *	ext4_find_shared - find the indirect blocks for partial truncation.
768 *	@inode:	  inode in question
769 *	@depth:	  depth of the affected branch
770 *	@offsets: offsets of pointers in that branch (see ext4_block_to_path)
771 *	@chain:	  place to store the pointers to partial indirect blocks
772 *	@top:	  place to the (detached) top of branch
773 *
774 *	This is a helper function used by ext4_truncate().
775 *
776 *	When we do truncate() we may have to clean the ends of several
777 *	indirect blocks but leave the blocks themselves alive. Block is
778 *	partially truncated if some data below the new i_size is referred
779 *	from it (and it is on the path to the first completely truncated
780 *	data block, indeed).  We have to free the top of that path along
781 *	with everything to the right of the path. Since no allocation
782 *	past the truncation point is possible until ext4_truncate()
783 *	finishes, we may safely do the latter, but top of branch may
784 *	require special attention - pageout below the truncation point
785 *	might try to populate it.
786 *
787 *	We atomically detach the top of branch from the tree, store the
788 *	block number of its root in *@top, pointers to buffer_heads of
789 *	partially truncated blocks - in @chain[].bh and pointers to
790 *	their last elements that should not be removed - in
791 *	@chain[].p. Return value is the pointer to last filled element
792 *	of @chain.
793 *
794 *	The work left to caller to do the actual freeing of subtrees:
795 *		a) free the subtree starting from *@top
796 *		b) free the subtrees whose roots are stored in
797 *			(@chain[i].p+1 .. end of @chain[i].bh->b_data)
798 *		c) free the subtrees growing from the inode past the @chain[0].
799 *			(no partially truncated stuff there).  */
800
801static Indirect *ext4_find_shared(struct inode *inode, int depth,
802				  ext4_lblk_t offsets[4], Indirect chain[4],
803				  __le32 *top)
804{
805	Indirect *partial, *p;
806	int k, err;
807
808	*top = 0;
809	/* Make k index the deepest non-null offset + 1 */
810	for (k = depth; k > 1 && !offsets[k-1]; k--)
811		;
812	partial = ext4_get_branch(inode, k, offsets, chain, &err);
813	/* Writer: pointers */
814	if (!partial)
815		partial = chain + k-1;
816	/*
817	 * If the branch acquired continuation since we've looked at it -
818	 * fine, it should all survive and (new) top doesn't belong to us.
819	 */
820	if (!partial->key && *partial->p)
821		/* Writer: end */
822		goto no_top;
823	for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
824		;
825	/*
826	 * OK, we've found the last block that must survive. The rest of our
827	 * branch should be detached before unlocking. However, if that rest
828	 * of branch is all ours and does not grow immediately from the inode
829	 * it's easier to cheat and just decrement partial->p.
830	 */
831	if (p == chain + k - 1 && p > chain) {
832		p->p--;
833	} else {
834		*top = *p->p;
835		/* Nope, don't do this in ext4.  Must leave the tree intact */
836#if 0
837		*p->p = 0;
838#endif
839	}
840	/* Writer: end */
841
842	while (partial > p) {
843		brelse(partial->bh);
844		partial--;
845	}
846no_top:
847	return partial;
848}
849
850/*
851 * Zero a number of block pointers in either an inode or an indirect block.
852 * If we restart the transaction we must again get write access to the
853 * indirect block for further modification.
854 *
855 * We release `count' blocks on disk, but (last - first) may be greater
856 * than `count' because there can be holes in there.
857 *
858 * Return 0 on success, 1 on invalid block range
859 * and < 0 on fatal error.
860 */
861static int ext4_clear_blocks(handle_t *handle, struct inode *inode,
862			     struct buffer_head *bh,
863			     ext4_fsblk_t block_to_free,
864			     unsigned long count, __le32 *first,
865			     __le32 *last)
866{
867	__le32 *p;
868	int	flags = EXT4_FREE_BLOCKS_VALIDATED;
869	int	err;
870
871	if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode) ||
872	    ext4_test_inode_flag(inode, EXT4_INODE_EA_INODE))
873		flags |= EXT4_FREE_BLOCKS_FORGET | EXT4_FREE_BLOCKS_METADATA;
874	else if (ext4_should_journal_data(inode))
875		flags |= EXT4_FREE_BLOCKS_FORGET;
876
877	if (!ext4_inode_block_valid(inode, block_to_free, count)) {
878		EXT4_ERROR_INODE(inode, "attempt to clear invalid "
879				 "blocks %llu len %lu",
880				 (unsigned long long) block_to_free, count);
881		return 1;
882	}
883
884	err = ext4_ind_truncate_ensure_credits(handle, inode, bh,
885				ext4_free_data_revoke_credits(inode, count));
886	if (err < 0)
887		goto out_err;
888
889	for (p = first; p < last; p++)
890		*p = 0;
891
892	ext4_free_blocks(handle, inode, NULL, block_to_free, count, flags);
893	return 0;
894out_err:
895	ext4_std_error(inode->i_sb, err);
896	return err;
897}
898
899/**
900 * ext4_free_data - free a list of data blocks
901 * @handle:	handle for this transaction
902 * @inode:	inode we are dealing with
903 * @this_bh:	indirect buffer_head which contains *@first and *@last
904 * @first:	array of block numbers
905 * @last:	points immediately past the end of array
906 *
907 * We are freeing all blocks referred from that array (numbers are stored as
908 * little-endian 32-bit) and updating @inode->i_blocks appropriately.
909 *
910 * We accumulate contiguous runs of blocks to free.  Conveniently, if these
911 * blocks are contiguous then releasing them at one time will only affect one
912 * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
913 * actually use a lot of journal space.
914 *
915 * @this_bh will be %NULL if @first and @last point into the inode's direct
916 * block pointers.
917 */
918static void ext4_free_data(handle_t *handle, struct inode *inode,
919			   struct buffer_head *this_bh,
920			   __le32 *first, __le32 *last)
921{
922	ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
923	unsigned long count = 0;	    /* Number of blocks in the run */
924	__le32 *block_to_free_p = NULL;	    /* Pointer into inode/ind
925					       corresponding to
926					       block_to_free */
927	ext4_fsblk_t nr;		    /* Current block # */
928	__le32 *p;			    /* Pointer into inode/ind
929					       for current block */
930	int err = 0;
931
932	if (this_bh) {				/* For indirect block */
933		BUFFER_TRACE(this_bh, "get_write_access");
934		err = ext4_journal_get_write_access(handle, this_bh);
935		/* Important: if we can't update the indirect pointers
936		 * to the blocks, we can't free them. */
937		if (err)
938			return;
939	}
940
941	for (p = first; p < last; p++) {
942		nr = le32_to_cpu(*p);
943		if (nr) {
944			/* accumulate blocks to free if they're contiguous */
945			if (count == 0) {
946				block_to_free = nr;
947				block_to_free_p = p;
948				count = 1;
949			} else if (nr == block_to_free + count) {
950				count++;
951			} else {
952				err = ext4_clear_blocks(handle, inode, this_bh,
953						        block_to_free, count,
954						        block_to_free_p, p);
955				if (err)
956					break;
957				block_to_free = nr;
958				block_to_free_p = p;
959				count = 1;
960			}
961		}
962	}
963
964	if (!err && count > 0)
965		err = ext4_clear_blocks(handle, inode, this_bh, block_to_free,
966					count, block_to_free_p, p);
967	if (err < 0)
968		/* fatal error */
969		return;
970
971	if (this_bh) {
972		BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
973
974		/*
975		 * The buffer head should have an attached journal head at this
976		 * point. However, if the data is corrupted and an indirect
977		 * block pointed to itself, it would have been detached when
978		 * the block was cleared. Check for this instead of OOPSing.
979		 */
980		if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
981			ext4_handle_dirty_metadata(handle, inode, this_bh);
982		else
983			EXT4_ERROR_INODE(inode,
984					 "circular indirect block detected at "
985					 "block %llu",
986				(unsigned long long) this_bh->b_blocknr);
987	}
988}
989
990/**
991 *	ext4_free_branches - free an array of branches
992 *	@handle: JBD handle for this transaction
993 *	@inode:	inode we are dealing with
994 *	@parent_bh: the buffer_head which contains *@first and *@last
995 *	@first:	array of block numbers
996 *	@last:	pointer immediately past the end of array
997 *	@depth:	depth of the branches to free
998 *
999 *	We are freeing all blocks referred from these branches (numbers are
1000 *	stored as little-endian 32-bit) and updating @inode->i_blocks
1001 *	appropriately.
1002 */
1003static void ext4_free_branches(handle_t *handle, struct inode *inode,
1004			       struct buffer_head *parent_bh,
1005			       __le32 *first, __le32 *last, int depth)
1006{
1007	ext4_fsblk_t nr;
1008	__le32 *p;
1009
1010	if (ext4_handle_is_aborted(handle))
1011		return;
1012
1013	if (depth--) {
1014		struct buffer_head *bh;
1015		int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1016		p = last;
1017		while (--p >= first) {
1018			nr = le32_to_cpu(*p);
1019			if (!nr)
1020				continue;		/* A hole */
1021
1022			if (!ext4_inode_block_valid(inode, nr, 1)) {
1023				EXT4_ERROR_INODE(inode,
1024						 "invalid indirect mapped "
1025						 "block %lu (level %d)",
1026						 (unsigned long) nr, depth);
1027				break;
1028			}
1029
1030			/* Go read the buffer for the next level down */
1031			bh = ext4_sb_bread(inode->i_sb, nr, 0);
1032
1033			/*
1034			 * A read failure? Report error and clear slot
1035			 * (should be rare).
1036			 */
1037			if (IS_ERR(bh)) {
1038				ext4_error_inode_block(inode, nr, -PTR_ERR(bh),
1039						       "Read failure");
1040				continue;
1041			}
1042
1043			/* This zaps the entire block.  Bottom up. */
1044			BUFFER_TRACE(bh, "free child branches");
1045			ext4_free_branches(handle, inode, bh,
1046					(__le32 *) bh->b_data,
1047					(__le32 *) bh->b_data + addr_per_block,
1048					depth);
1049			brelse(bh);
1050
1051			/*
1052			 * Everything below this pointer has been
1053			 * released.  Now let this top-of-subtree go.
1054			 *
1055			 * We want the freeing of this indirect block to be
1056			 * atomic in the journal with the updating of the
1057			 * bitmap block which owns it.  So make some room in
1058			 * the journal.
1059			 *
1060			 * We zero the parent pointer *after* freeing its
1061			 * pointee in the bitmaps, so if extend_transaction()
1062			 * for some reason fails to put the bitmap changes and
1063			 * the release into the same transaction, recovery
1064			 * will merely complain about releasing a free block,
1065			 * rather than leaking blocks.
1066			 */
1067			if (ext4_handle_is_aborted(handle))
1068				return;
1069			if (ext4_ind_truncate_ensure_credits(handle, inode,
1070					NULL,
1071					ext4_free_metadata_revoke_credits(
1072							inode->i_sb, 1)) < 0)
1073				return;
1074
1075			/*
1076			 * The forget flag here is critical because if
1077			 * we are journaling (and not doing data
1078			 * journaling), we have to make sure a revoke
1079			 * record is written to prevent the journal
1080			 * replay from overwriting the (former)
1081			 * indirect block if it gets reallocated as a
1082			 * data block.  This must happen in the same
1083			 * transaction where the data blocks are
1084			 * actually freed.
1085			 */
1086			ext4_free_blocks(handle, inode, NULL, nr, 1,
1087					 EXT4_FREE_BLOCKS_METADATA|
1088					 EXT4_FREE_BLOCKS_FORGET);
1089
1090			if (parent_bh) {
1091				/*
1092				 * The block which we have just freed is
1093				 * pointed to by an indirect block: journal it
1094				 */
1095				BUFFER_TRACE(parent_bh, "get_write_access");
1096				if (!ext4_journal_get_write_access(handle,
1097								   parent_bh)){
1098					*p = 0;
1099					BUFFER_TRACE(parent_bh,
1100					"call ext4_handle_dirty_metadata");
1101					ext4_handle_dirty_metadata(handle,
1102								   inode,
1103								   parent_bh);
1104				}
1105			}
1106		}
1107	} else {
1108		/* We have reached the bottom of the tree. */
1109		BUFFER_TRACE(parent_bh, "free data blocks");
1110		ext4_free_data(handle, inode, parent_bh, first, last);
1111	}
1112}
1113
1114void ext4_ind_truncate(handle_t *handle, struct inode *inode)
1115{
1116	struct ext4_inode_info *ei = EXT4_I(inode);
1117	__le32 *i_data = ei->i_data;
1118	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1119	ext4_lblk_t offsets[4];
1120	Indirect chain[4];
1121	Indirect *partial;
1122	__le32 nr = 0;
1123	int n = 0;
1124	ext4_lblk_t last_block, max_block;
1125	unsigned blocksize = inode->i_sb->s_blocksize;
1126
1127	last_block = (inode->i_size + blocksize-1)
1128					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1129	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1130					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1131
1132	if (last_block != max_block) {
1133		n = ext4_block_to_path(inode, last_block, offsets, NULL);
1134		if (n == 0)
1135			return;
1136	}
1137
1138	ext4_es_remove_extent(inode, last_block, EXT_MAX_BLOCKS - last_block);
1139
1140	/*
1141	 * The orphan list entry will now protect us from any crash which
1142	 * occurs before the truncate completes, so it is now safe to propagate
1143	 * the new, shorter inode size (held for now in i_size) into the
1144	 * on-disk inode. We do this via i_disksize, which is the value which
1145	 * ext4 *really* writes onto the disk inode.
1146	 */
1147	ei->i_disksize = inode->i_size;
1148
1149	if (last_block == max_block) {
1150		/*
1151		 * It is unnecessary to free any data blocks if last_block is
1152		 * equal to the indirect block limit.
1153		 */
1154		return;
1155	} else if (n == 1) {		/* direct blocks */
1156		ext4_free_data(handle, inode, NULL, i_data+offsets[0],
1157			       i_data + EXT4_NDIR_BLOCKS);
1158		goto do_indirects;
1159	}
1160
1161	partial = ext4_find_shared(inode, n, offsets, chain, &nr);
1162	/* Kill the top of shared branch (not detached) */
1163	if (nr) {
1164		if (partial == chain) {
1165			/* Shared branch grows from the inode */
1166			ext4_free_branches(handle, inode, NULL,
1167					   &nr, &nr+1, (chain+n-1) - partial);
1168			*partial->p = 0;
1169			/*
1170			 * We mark the inode dirty prior to restart,
1171			 * and prior to stop.  No need for it here.
1172			 */
1173		} else {
1174			/* Shared branch grows from an indirect block */
1175			BUFFER_TRACE(partial->bh, "get_write_access");
1176			ext4_free_branches(handle, inode, partial->bh,
1177					partial->p,
1178					partial->p+1, (chain+n-1) - partial);
1179		}
1180	}
1181	/* Clear the ends of indirect blocks on the shared branch */
1182	while (partial > chain) {
1183		ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
1184				   (__le32*)partial->bh->b_data+addr_per_block,
1185				   (chain+n-1) - partial);
1186		BUFFER_TRACE(partial->bh, "call brelse");
1187		brelse(partial->bh);
1188		partial--;
1189	}
1190do_indirects:
1191	/* Kill the remaining (whole) subtrees */
1192	switch (offsets[0]) {
1193	default:
1194		nr = i_data[EXT4_IND_BLOCK];
1195		if (nr) {
1196			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1197			i_data[EXT4_IND_BLOCK] = 0;
1198		}
1199		fallthrough;
1200	case EXT4_IND_BLOCK:
1201		nr = i_data[EXT4_DIND_BLOCK];
1202		if (nr) {
1203			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1204			i_data[EXT4_DIND_BLOCK] = 0;
1205		}
1206		fallthrough;
1207	case EXT4_DIND_BLOCK:
1208		nr = i_data[EXT4_TIND_BLOCK];
1209		if (nr) {
1210			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1211			i_data[EXT4_TIND_BLOCK] = 0;
1212		}
1213		fallthrough;
1214	case EXT4_TIND_BLOCK:
1215		;
1216	}
1217}
1218
1219/**
1220 *	ext4_ind_remove_space - remove space from the range
1221 *	@handle: JBD handle for this transaction
1222 *	@inode:	inode we are dealing with
1223 *	@start:	First block to remove
1224 *	@end:	One block after the last block to remove (exclusive)
1225 *
1226 *	Free the blocks in the defined range (end is exclusive endpoint of
1227 *	range). This is used by ext4_punch_hole().
1228 */
1229int ext4_ind_remove_space(handle_t *handle, struct inode *inode,
1230			  ext4_lblk_t start, ext4_lblk_t end)
1231{
1232	struct ext4_inode_info *ei = EXT4_I(inode);
1233	__le32 *i_data = ei->i_data;
1234	int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1235	ext4_lblk_t offsets[4], offsets2[4];
1236	Indirect chain[4], chain2[4];
1237	Indirect *partial, *partial2;
1238	Indirect *p = NULL, *p2 = NULL;
1239	ext4_lblk_t max_block;
1240	__le32 nr = 0, nr2 = 0;
1241	int n = 0, n2 = 0;
1242	unsigned blocksize = inode->i_sb->s_blocksize;
1243
1244	max_block = (EXT4_SB(inode->i_sb)->s_bitmap_maxbytes + blocksize-1)
1245					>> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
1246	if (end >= max_block)
1247		end = max_block;
1248	if ((start >= end) || (start > max_block))
1249		return 0;
1250
1251	n = ext4_block_to_path(inode, start, offsets, NULL);
1252	n2 = ext4_block_to_path(inode, end, offsets2, NULL);
1253
1254	BUG_ON(n > n2);
1255
1256	if ((n == 1) && (n == n2)) {
1257		/* We're punching only within direct block range */
1258		ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1259			       i_data + offsets2[0]);
1260		return 0;
1261	} else if (n2 > n) {
1262		/*
1263		 * Start and end are on a different levels so we're going to
1264		 * free partial block at start, and partial block at end of
1265		 * the range. If there are some levels in between then
1266		 * do_indirects label will take care of that.
1267		 */
1268
1269		if (n == 1) {
1270			/*
1271			 * Start is at the direct block level, free
1272			 * everything to the end of the level.
1273			 */
1274			ext4_free_data(handle, inode, NULL, i_data + offsets[0],
1275				       i_data + EXT4_NDIR_BLOCKS);
1276			goto end_range;
1277		}
1278
1279
1280		partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1281		if (nr) {
1282			if (partial == chain) {
1283				/* Shared branch grows from the inode */
1284				ext4_free_branches(handle, inode, NULL,
1285					   &nr, &nr+1, (chain+n-1) - partial);
1286				*partial->p = 0;
1287			} else {
1288				/* Shared branch grows from an indirect block */
1289				BUFFER_TRACE(partial->bh, "get_write_access");
1290				ext4_free_branches(handle, inode, partial->bh,
1291					partial->p,
1292					partial->p+1, (chain+n-1) - partial);
1293			}
1294		}
1295
1296		/*
1297		 * Clear the ends of indirect blocks on the shared branch
1298		 * at the start of the range
1299		 */
1300		while (partial > chain) {
1301			ext4_free_branches(handle, inode, partial->bh,
1302				partial->p + 1,
1303				(__le32 *)partial->bh->b_data+addr_per_block,
1304				(chain+n-1) - partial);
1305			partial--;
1306		}
1307
1308end_range:
1309		partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1310		if (nr2) {
1311			if (partial2 == chain2) {
1312				/*
1313				 * Remember, end is exclusive so here we're at
1314				 * the start of the next level we're not going
1315				 * to free. Everything was covered by the start
1316				 * of the range.
1317				 */
1318				goto do_indirects;
1319			}
1320		} else {
1321			/*
1322			 * ext4_find_shared returns Indirect structure which
1323			 * points to the last element which should not be
1324			 * removed by truncate. But this is end of the range
1325			 * in punch_hole so we need to point to the next element
1326			 */
1327			partial2->p++;
1328		}
1329
1330		/*
1331		 * Clear the ends of indirect blocks on the shared branch
1332		 * at the end of the range
1333		 */
1334		while (partial2 > chain2) {
1335			ext4_free_branches(handle, inode, partial2->bh,
1336					   (__le32 *)partial2->bh->b_data,
1337					   partial2->p,
1338					   (chain2+n2-1) - partial2);
1339			partial2--;
1340		}
1341		goto do_indirects;
1342	}
1343
1344	/* Punch happened within the same level (n == n2) */
1345	partial = p = ext4_find_shared(inode, n, offsets, chain, &nr);
1346	partial2 = p2 = ext4_find_shared(inode, n2, offsets2, chain2, &nr2);
1347
1348	/* Free top, but only if partial2 isn't its subtree. */
1349	if (nr) {
1350		int level = min(partial - chain, partial2 - chain2);
1351		int i;
1352		int subtree = 1;
1353
1354		for (i = 0; i <= level; i++) {
1355			if (offsets[i] != offsets2[i]) {
1356				subtree = 0;
1357				break;
1358			}
1359		}
1360
1361		if (!subtree) {
1362			if (partial == chain) {
1363				/* Shared branch grows from the inode */
1364				ext4_free_branches(handle, inode, NULL,
1365						   &nr, &nr+1,
1366						   (chain+n-1) - partial);
1367				*partial->p = 0;
1368			} else {
1369				/* Shared branch grows from an indirect block */
1370				BUFFER_TRACE(partial->bh, "get_write_access");
1371				ext4_free_branches(handle, inode, partial->bh,
1372						   partial->p,
1373						   partial->p+1,
1374						   (chain+n-1) - partial);
1375			}
1376		}
1377	}
1378
1379	if (!nr2) {
1380		/*
1381		 * ext4_find_shared returns Indirect structure which
1382		 * points to the last element which should not be
1383		 * removed by truncate. But this is end of the range
1384		 * in punch_hole so we need to point to the next element
1385		 */
1386		partial2->p++;
1387	}
1388
1389	while (partial > chain || partial2 > chain2) {
1390		int depth = (chain+n-1) - partial;
1391		int depth2 = (chain2+n2-1) - partial2;
1392
1393		if (partial > chain && partial2 > chain2 &&
1394		    partial->bh->b_blocknr == partial2->bh->b_blocknr) {
1395			/*
1396			 * We've converged on the same block. Clear the range,
1397			 * then we're done.
1398			 */
1399			ext4_free_branches(handle, inode, partial->bh,
1400					   partial->p + 1,
1401					   partial2->p,
1402					   (chain+n-1) - partial);
1403			goto cleanup;
1404		}
1405
1406		/*
1407		 * The start and end partial branches may not be at the same
1408		 * level even though the punch happened within one level. So, we
1409		 * give them a chance to arrive at the same level, then walk
1410		 * them in step with each other until we converge on the same
1411		 * block.
1412		 */
1413		if (partial > chain && depth <= depth2) {
1414			ext4_free_branches(handle, inode, partial->bh,
1415					   partial->p + 1,
1416					   (__le32 *)partial->bh->b_data+addr_per_block,
1417					   (chain+n-1) - partial);
1418			partial--;
1419		}
1420		if (partial2 > chain2 && depth2 <= depth) {
1421			ext4_free_branches(handle, inode, partial2->bh,
1422					   (__le32 *)partial2->bh->b_data,
1423					   partial2->p,
1424					   (chain2+n2-1) - partial2);
1425			partial2--;
1426		}
1427	}
1428
1429cleanup:
1430	while (p && p > chain) {
1431		BUFFER_TRACE(p->bh, "call brelse");
1432		brelse(p->bh);
1433		p--;
1434	}
1435	while (p2 && p2 > chain2) {
1436		BUFFER_TRACE(p2->bh, "call brelse");
1437		brelse(p2->bh);
1438		p2--;
1439	}
1440	return 0;
1441
1442do_indirects:
1443	/* Kill the remaining (whole) subtrees */
1444	switch (offsets[0]) {
1445	default:
1446		if (++n >= n2)
1447			break;
1448		nr = i_data[EXT4_IND_BLOCK];
1449		if (nr) {
1450			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
1451			i_data[EXT4_IND_BLOCK] = 0;
1452		}
1453		fallthrough;
1454	case EXT4_IND_BLOCK:
1455		if (++n >= n2)
1456			break;
1457		nr = i_data[EXT4_DIND_BLOCK];
1458		if (nr) {
1459			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
1460			i_data[EXT4_DIND_BLOCK] = 0;
1461		}
1462		fallthrough;
1463	case EXT4_DIND_BLOCK:
1464		if (++n >= n2)
1465			break;
1466		nr = i_data[EXT4_TIND_BLOCK];
1467		if (nr) {
1468			ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
1469			i_data[EXT4_TIND_BLOCK] = 0;
1470		}
1471		fallthrough;
1472	case EXT4_TIND_BLOCK:
1473		;
1474	}
1475	goto cleanup;
1476}
1477