xref: /openbmc/linux/fs/ext4/extents.c (revision 6774def6)
1 /*
2  * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
3  * Written by Alex Tomas <alex@clusterfs.com>
4  *
5  * Architecture independence:
6  *   Copyright (c) 2005, Bull S.A.
7  *   Written by Pierre Peiffer <pierre.peiffer@bull.net>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public Licens
19  * along with this program; if not, write to the Free Software
20  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
21  */
22 
23 /*
24  * Extents support for EXT4
25  *
26  * TODO:
27  *   - ext4*_error() should be used in some situations
28  *   - analyze all BUG()/BUG_ON(), use -EIO where appropriate
29  *   - smart tree reduction
30  */
31 
32 #include <linux/fs.h>
33 #include <linux/time.h>
34 #include <linux/jbd2.h>
35 #include <linux/highuid.h>
36 #include <linux/pagemap.h>
37 #include <linux/quotaops.h>
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include <asm/uaccess.h>
41 #include <linux/fiemap.h>
42 #include "ext4_jbd2.h"
43 #include "ext4_extents.h"
44 #include "xattr.h"
45 
46 #include <trace/events/ext4.h>
47 
48 /*
49  * used by extent splitting.
50  */
51 #define EXT4_EXT_MAY_ZEROOUT	0x1  /* safe to zeroout if split fails \
52 					due to ENOSPC */
53 #define EXT4_EXT_MARK_UNWRIT1	0x2  /* mark first half unwritten */
54 #define EXT4_EXT_MARK_UNWRIT2	0x4  /* mark second half unwritten */
55 
56 #define EXT4_EXT_DATA_VALID1	0x8  /* first half contains valid data */
57 #define EXT4_EXT_DATA_VALID2	0x10 /* second half contains valid data */
58 
59 static __le32 ext4_extent_block_csum(struct inode *inode,
60 				     struct ext4_extent_header *eh)
61 {
62 	struct ext4_inode_info *ei = EXT4_I(inode);
63 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
64 	__u32 csum;
65 
66 	csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)eh,
67 			   EXT4_EXTENT_TAIL_OFFSET(eh));
68 	return cpu_to_le32(csum);
69 }
70 
71 static int ext4_extent_block_csum_verify(struct inode *inode,
72 					 struct ext4_extent_header *eh)
73 {
74 	struct ext4_extent_tail *et;
75 
76 	if (!ext4_has_metadata_csum(inode->i_sb))
77 		return 1;
78 
79 	et = find_ext4_extent_tail(eh);
80 	if (et->et_checksum != ext4_extent_block_csum(inode, eh))
81 		return 0;
82 	return 1;
83 }
84 
85 static void ext4_extent_block_csum_set(struct inode *inode,
86 				       struct ext4_extent_header *eh)
87 {
88 	struct ext4_extent_tail *et;
89 
90 	if (!ext4_has_metadata_csum(inode->i_sb))
91 		return;
92 
93 	et = find_ext4_extent_tail(eh);
94 	et->et_checksum = ext4_extent_block_csum(inode, eh);
95 }
96 
97 static int ext4_split_extent(handle_t *handle,
98 				struct inode *inode,
99 				struct ext4_ext_path **ppath,
100 				struct ext4_map_blocks *map,
101 				int split_flag,
102 				int flags);
103 
104 static int ext4_split_extent_at(handle_t *handle,
105 			     struct inode *inode,
106 			     struct ext4_ext_path **ppath,
107 			     ext4_lblk_t split,
108 			     int split_flag,
109 			     int flags);
110 
111 static int ext4_find_delayed_extent(struct inode *inode,
112 				    struct extent_status *newes);
113 
114 static int ext4_ext_truncate_extend_restart(handle_t *handle,
115 					    struct inode *inode,
116 					    int needed)
117 {
118 	int err;
119 
120 	if (!ext4_handle_valid(handle))
121 		return 0;
122 	if (handle->h_buffer_credits > needed)
123 		return 0;
124 	err = ext4_journal_extend(handle, needed);
125 	if (err <= 0)
126 		return err;
127 	err = ext4_truncate_restart_trans(handle, inode, needed);
128 	if (err == 0)
129 		err = -EAGAIN;
130 
131 	return err;
132 }
133 
134 /*
135  * could return:
136  *  - EROFS
137  *  - ENOMEM
138  */
139 static int ext4_ext_get_access(handle_t *handle, struct inode *inode,
140 				struct ext4_ext_path *path)
141 {
142 	if (path->p_bh) {
143 		/* path points to block */
144 		BUFFER_TRACE(path->p_bh, "get_write_access");
145 		return ext4_journal_get_write_access(handle, path->p_bh);
146 	}
147 	/* path points to leaf/index in inode body */
148 	/* we use in-core data, no need to protect them */
149 	return 0;
150 }
151 
152 /*
153  * could return:
154  *  - EROFS
155  *  - ENOMEM
156  *  - EIO
157  */
158 int __ext4_ext_dirty(const char *where, unsigned int line, handle_t *handle,
159 		     struct inode *inode, struct ext4_ext_path *path)
160 {
161 	int err;
162 
163 	WARN_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
164 	if (path->p_bh) {
165 		ext4_extent_block_csum_set(inode, ext_block_hdr(path->p_bh));
166 		/* path points to block */
167 		err = __ext4_handle_dirty_metadata(where, line, handle,
168 						   inode, path->p_bh);
169 	} else {
170 		/* path points to leaf/index in inode body */
171 		err = ext4_mark_inode_dirty(handle, inode);
172 	}
173 	return err;
174 }
175 
176 static ext4_fsblk_t ext4_ext_find_goal(struct inode *inode,
177 			      struct ext4_ext_path *path,
178 			      ext4_lblk_t block)
179 {
180 	if (path) {
181 		int depth = path->p_depth;
182 		struct ext4_extent *ex;
183 
184 		/*
185 		 * Try to predict block placement assuming that we are
186 		 * filling in a file which will eventually be
187 		 * non-sparse --- i.e., in the case of libbfd writing
188 		 * an ELF object sections out-of-order but in a way
189 		 * the eventually results in a contiguous object or
190 		 * executable file, or some database extending a table
191 		 * space file.  However, this is actually somewhat
192 		 * non-ideal if we are writing a sparse file such as
193 		 * qemu or KVM writing a raw image file that is going
194 		 * to stay fairly sparse, since it will end up
195 		 * fragmenting the file system's free space.  Maybe we
196 		 * should have some hueristics or some way to allow
197 		 * userspace to pass a hint to file system,
198 		 * especially if the latter case turns out to be
199 		 * common.
200 		 */
201 		ex = path[depth].p_ext;
202 		if (ex) {
203 			ext4_fsblk_t ext_pblk = ext4_ext_pblock(ex);
204 			ext4_lblk_t ext_block = le32_to_cpu(ex->ee_block);
205 
206 			if (block > ext_block)
207 				return ext_pblk + (block - ext_block);
208 			else
209 				return ext_pblk - (ext_block - block);
210 		}
211 
212 		/* it looks like index is empty;
213 		 * try to find starting block from index itself */
214 		if (path[depth].p_bh)
215 			return path[depth].p_bh->b_blocknr;
216 	}
217 
218 	/* OK. use inode's group */
219 	return ext4_inode_to_goal_block(inode);
220 }
221 
222 /*
223  * Allocation for a meta data block
224  */
225 static ext4_fsblk_t
226 ext4_ext_new_meta_block(handle_t *handle, struct inode *inode,
227 			struct ext4_ext_path *path,
228 			struct ext4_extent *ex, int *err, unsigned int flags)
229 {
230 	ext4_fsblk_t goal, newblock;
231 
232 	goal = ext4_ext_find_goal(inode, path, le32_to_cpu(ex->ee_block));
233 	newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
234 					NULL, err);
235 	return newblock;
236 }
237 
238 static inline int ext4_ext_space_block(struct inode *inode, int check)
239 {
240 	int size;
241 
242 	size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
243 			/ sizeof(struct ext4_extent);
244 #ifdef AGGRESSIVE_TEST
245 	if (!check && size > 6)
246 		size = 6;
247 #endif
248 	return size;
249 }
250 
251 static inline int ext4_ext_space_block_idx(struct inode *inode, int check)
252 {
253 	int size;
254 
255 	size = (inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
256 			/ sizeof(struct ext4_extent_idx);
257 #ifdef AGGRESSIVE_TEST
258 	if (!check && size > 5)
259 		size = 5;
260 #endif
261 	return size;
262 }
263 
264 static inline int ext4_ext_space_root(struct inode *inode, int check)
265 {
266 	int size;
267 
268 	size = sizeof(EXT4_I(inode)->i_data);
269 	size -= sizeof(struct ext4_extent_header);
270 	size /= sizeof(struct ext4_extent);
271 #ifdef AGGRESSIVE_TEST
272 	if (!check && size > 3)
273 		size = 3;
274 #endif
275 	return size;
276 }
277 
278 static inline int ext4_ext_space_root_idx(struct inode *inode, int check)
279 {
280 	int size;
281 
282 	size = sizeof(EXT4_I(inode)->i_data);
283 	size -= sizeof(struct ext4_extent_header);
284 	size /= sizeof(struct ext4_extent_idx);
285 #ifdef AGGRESSIVE_TEST
286 	if (!check && size > 4)
287 		size = 4;
288 #endif
289 	return size;
290 }
291 
292 static inline int
293 ext4_force_split_extent_at(handle_t *handle, struct inode *inode,
294 			   struct ext4_ext_path **ppath, ext4_lblk_t lblk,
295 			   int nofail)
296 {
297 	struct ext4_ext_path *path = *ppath;
298 	int unwritten = ext4_ext_is_unwritten(path[path->p_depth].p_ext);
299 
300 	return ext4_split_extent_at(handle, inode, ppath, lblk, unwritten ?
301 			EXT4_EXT_MARK_UNWRIT1|EXT4_EXT_MARK_UNWRIT2 : 0,
302 			EXT4_EX_NOCACHE | EXT4_GET_BLOCKS_PRE_IO |
303 			(nofail ? EXT4_GET_BLOCKS_METADATA_NOFAIL:0));
304 }
305 
306 /*
307  * Calculate the number of metadata blocks needed
308  * to allocate @blocks
309  * Worse case is one block per extent
310  */
311 int ext4_ext_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
312 {
313 	struct ext4_inode_info *ei = EXT4_I(inode);
314 	int idxs;
315 
316 	idxs = ((inode->i_sb->s_blocksize - sizeof(struct ext4_extent_header))
317 		/ sizeof(struct ext4_extent_idx));
318 
319 	/*
320 	 * If the new delayed allocation block is contiguous with the
321 	 * previous da block, it can share index blocks with the
322 	 * previous block, so we only need to allocate a new index
323 	 * block every idxs leaf blocks.  At ldxs**2 blocks, we need
324 	 * an additional index block, and at ldxs**3 blocks, yet
325 	 * another index blocks.
326 	 */
327 	if (ei->i_da_metadata_calc_len &&
328 	    ei->i_da_metadata_calc_last_lblock+1 == lblock) {
329 		int num = 0;
330 
331 		if ((ei->i_da_metadata_calc_len % idxs) == 0)
332 			num++;
333 		if ((ei->i_da_metadata_calc_len % (idxs*idxs)) == 0)
334 			num++;
335 		if ((ei->i_da_metadata_calc_len % (idxs*idxs*idxs)) == 0) {
336 			num++;
337 			ei->i_da_metadata_calc_len = 0;
338 		} else
339 			ei->i_da_metadata_calc_len++;
340 		ei->i_da_metadata_calc_last_lblock++;
341 		return num;
342 	}
343 
344 	/*
345 	 * In the worst case we need a new set of index blocks at
346 	 * every level of the inode's extent tree.
347 	 */
348 	ei->i_da_metadata_calc_len = 1;
349 	ei->i_da_metadata_calc_last_lblock = lblock;
350 	return ext_depth(inode) + 1;
351 }
352 
353 static int
354 ext4_ext_max_entries(struct inode *inode, int depth)
355 {
356 	int max;
357 
358 	if (depth == ext_depth(inode)) {
359 		if (depth == 0)
360 			max = ext4_ext_space_root(inode, 1);
361 		else
362 			max = ext4_ext_space_root_idx(inode, 1);
363 	} else {
364 		if (depth == 0)
365 			max = ext4_ext_space_block(inode, 1);
366 		else
367 			max = ext4_ext_space_block_idx(inode, 1);
368 	}
369 
370 	return max;
371 }
372 
373 static int ext4_valid_extent(struct inode *inode, struct ext4_extent *ext)
374 {
375 	ext4_fsblk_t block = ext4_ext_pblock(ext);
376 	int len = ext4_ext_get_actual_len(ext);
377 	ext4_lblk_t lblock = le32_to_cpu(ext->ee_block);
378 	ext4_lblk_t last = lblock + len - 1;
379 
380 	if (lblock > last)
381 		return 0;
382 	return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, len);
383 }
384 
385 static int ext4_valid_extent_idx(struct inode *inode,
386 				struct ext4_extent_idx *ext_idx)
387 {
388 	ext4_fsblk_t block = ext4_idx_pblock(ext_idx);
389 
390 	return ext4_data_block_valid(EXT4_SB(inode->i_sb), block, 1);
391 }
392 
393 static int ext4_valid_extent_entries(struct inode *inode,
394 				struct ext4_extent_header *eh,
395 				int depth)
396 {
397 	unsigned short entries;
398 	if (eh->eh_entries == 0)
399 		return 1;
400 
401 	entries = le16_to_cpu(eh->eh_entries);
402 
403 	if (depth == 0) {
404 		/* leaf entries */
405 		struct ext4_extent *ext = EXT_FIRST_EXTENT(eh);
406 		struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
407 		ext4_fsblk_t pblock = 0;
408 		ext4_lblk_t lblock = 0;
409 		ext4_lblk_t prev = 0;
410 		int len = 0;
411 		while (entries) {
412 			if (!ext4_valid_extent(inode, ext))
413 				return 0;
414 
415 			/* Check for overlapping extents */
416 			lblock = le32_to_cpu(ext->ee_block);
417 			len = ext4_ext_get_actual_len(ext);
418 			if ((lblock <= prev) && prev) {
419 				pblock = ext4_ext_pblock(ext);
420 				es->s_last_error_block = cpu_to_le64(pblock);
421 				return 0;
422 			}
423 			ext++;
424 			entries--;
425 			prev = lblock + len - 1;
426 		}
427 	} else {
428 		struct ext4_extent_idx *ext_idx = EXT_FIRST_INDEX(eh);
429 		while (entries) {
430 			if (!ext4_valid_extent_idx(inode, ext_idx))
431 				return 0;
432 			ext_idx++;
433 			entries--;
434 		}
435 	}
436 	return 1;
437 }
438 
439 static int __ext4_ext_check(const char *function, unsigned int line,
440 			    struct inode *inode, struct ext4_extent_header *eh,
441 			    int depth, ext4_fsblk_t pblk)
442 {
443 	const char *error_msg;
444 	int max = 0;
445 
446 	if (unlikely(eh->eh_magic != EXT4_EXT_MAGIC)) {
447 		error_msg = "invalid magic";
448 		goto corrupted;
449 	}
450 	if (unlikely(le16_to_cpu(eh->eh_depth) != depth)) {
451 		error_msg = "unexpected eh_depth";
452 		goto corrupted;
453 	}
454 	if (unlikely(eh->eh_max == 0)) {
455 		error_msg = "invalid eh_max";
456 		goto corrupted;
457 	}
458 	max = ext4_ext_max_entries(inode, depth);
459 	if (unlikely(le16_to_cpu(eh->eh_max) > max)) {
460 		error_msg = "too large eh_max";
461 		goto corrupted;
462 	}
463 	if (unlikely(le16_to_cpu(eh->eh_entries) > le16_to_cpu(eh->eh_max))) {
464 		error_msg = "invalid eh_entries";
465 		goto corrupted;
466 	}
467 	if (!ext4_valid_extent_entries(inode, eh, depth)) {
468 		error_msg = "invalid extent entries";
469 		goto corrupted;
470 	}
471 	/* Verify checksum on non-root extent tree nodes */
472 	if (ext_depth(inode) != depth &&
473 	    !ext4_extent_block_csum_verify(inode, eh)) {
474 		error_msg = "extent tree corrupted";
475 		goto corrupted;
476 	}
477 	return 0;
478 
479 corrupted:
480 	ext4_error_inode(inode, function, line, 0,
481 			 "pblk %llu bad header/extent: %s - magic %x, "
482 			 "entries %u, max %u(%u), depth %u(%u)",
483 			 (unsigned long long) pblk, error_msg,
484 			 le16_to_cpu(eh->eh_magic),
485 			 le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max),
486 			 max, le16_to_cpu(eh->eh_depth), depth);
487 	return -EIO;
488 }
489 
490 #define ext4_ext_check(inode, eh, depth, pblk)			\
491 	__ext4_ext_check(__func__, __LINE__, (inode), (eh), (depth), (pblk))
492 
493 int ext4_ext_check_inode(struct inode *inode)
494 {
495 	return ext4_ext_check(inode, ext_inode_hdr(inode), ext_depth(inode), 0);
496 }
497 
498 static struct buffer_head *
499 __read_extent_tree_block(const char *function, unsigned int line,
500 			 struct inode *inode, ext4_fsblk_t pblk, int depth,
501 			 int flags)
502 {
503 	struct buffer_head		*bh;
504 	int				err;
505 
506 	bh = sb_getblk(inode->i_sb, pblk);
507 	if (unlikely(!bh))
508 		return ERR_PTR(-ENOMEM);
509 
510 	if (!bh_uptodate_or_lock(bh)) {
511 		trace_ext4_ext_load_extent(inode, pblk, _RET_IP_);
512 		err = bh_submit_read(bh);
513 		if (err < 0)
514 			goto errout;
515 	}
516 	if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
517 		return bh;
518 	err = __ext4_ext_check(function, line, inode,
519 			       ext_block_hdr(bh), depth, pblk);
520 	if (err)
521 		goto errout;
522 	set_buffer_verified(bh);
523 	/*
524 	 * If this is a leaf block, cache all of its entries
525 	 */
526 	if (!(flags & EXT4_EX_NOCACHE) && depth == 0) {
527 		struct ext4_extent_header *eh = ext_block_hdr(bh);
528 		struct ext4_extent *ex = EXT_FIRST_EXTENT(eh);
529 		ext4_lblk_t prev = 0;
530 		int i;
531 
532 		for (i = le16_to_cpu(eh->eh_entries); i > 0; i--, ex++) {
533 			unsigned int status = EXTENT_STATUS_WRITTEN;
534 			ext4_lblk_t lblk = le32_to_cpu(ex->ee_block);
535 			int len = ext4_ext_get_actual_len(ex);
536 
537 			if (prev && (prev != lblk))
538 				ext4_es_cache_extent(inode, prev,
539 						     lblk - prev, ~0,
540 						     EXTENT_STATUS_HOLE);
541 
542 			if (ext4_ext_is_unwritten(ex))
543 				status = EXTENT_STATUS_UNWRITTEN;
544 			ext4_es_cache_extent(inode, lblk, len,
545 					     ext4_ext_pblock(ex), status);
546 			prev = lblk + len;
547 		}
548 	}
549 	return bh;
550 errout:
551 	put_bh(bh);
552 	return ERR_PTR(err);
553 
554 }
555 
556 #define read_extent_tree_block(inode, pblk, depth, flags)		\
557 	__read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
558 				 (depth), (flags))
559 
560 /*
561  * This function is called to cache a file's extent information in the
562  * extent status tree
563  */
564 int ext4_ext_precache(struct inode *inode)
565 {
566 	struct ext4_inode_info *ei = EXT4_I(inode);
567 	struct ext4_ext_path *path = NULL;
568 	struct buffer_head *bh;
569 	int i = 0, depth, ret = 0;
570 
571 	if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
572 		return 0;	/* not an extent-mapped inode */
573 
574 	down_read(&ei->i_data_sem);
575 	depth = ext_depth(inode);
576 
577 	path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
578 		       GFP_NOFS);
579 	if (path == NULL) {
580 		up_read(&ei->i_data_sem);
581 		return -ENOMEM;
582 	}
583 
584 	/* Don't cache anything if there are no external extent blocks */
585 	if (depth == 0)
586 		goto out;
587 	path[0].p_hdr = ext_inode_hdr(inode);
588 	ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
589 	if (ret)
590 		goto out;
591 	path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
592 	while (i >= 0) {
593 		/*
594 		 * If this is a leaf block or we've reached the end of
595 		 * the index block, go up
596 		 */
597 		if ((i == depth) ||
598 		    path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
599 			brelse(path[i].p_bh);
600 			path[i].p_bh = NULL;
601 			i--;
602 			continue;
603 		}
604 		bh = read_extent_tree_block(inode,
605 					    ext4_idx_pblock(path[i].p_idx++),
606 					    depth - i - 1,
607 					    EXT4_EX_FORCE_CACHE);
608 		if (IS_ERR(bh)) {
609 			ret = PTR_ERR(bh);
610 			break;
611 		}
612 		i++;
613 		path[i].p_bh = bh;
614 		path[i].p_hdr = ext_block_hdr(bh);
615 		path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
616 	}
617 	ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
618 out:
619 	up_read(&ei->i_data_sem);
620 	ext4_ext_drop_refs(path);
621 	kfree(path);
622 	return ret;
623 }
624 
625 #ifdef EXT_DEBUG
626 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
627 {
628 	int k, l = path->p_depth;
629 
630 	ext_debug("path:");
631 	for (k = 0; k <= l; k++, path++) {
632 		if (path->p_idx) {
633 		  ext_debug("  %d->%llu", le32_to_cpu(path->p_idx->ei_block),
634 			    ext4_idx_pblock(path->p_idx));
635 		} else if (path->p_ext) {
636 			ext_debug("  %d:[%d]%d:%llu ",
637 				  le32_to_cpu(path->p_ext->ee_block),
638 				  ext4_ext_is_unwritten(path->p_ext),
639 				  ext4_ext_get_actual_len(path->p_ext),
640 				  ext4_ext_pblock(path->p_ext));
641 		} else
642 			ext_debug("  []");
643 	}
644 	ext_debug("\n");
645 }
646 
647 static void ext4_ext_show_leaf(struct inode *inode, struct ext4_ext_path *path)
648 {
649 	int depth = ext_depth(inode);
650 	struct ext4_extent_header *eh;
651 	struct ext4_extent *ex;
652 	int i;
653 
654 	if (!path)
655 		return;
656 
657 	eh = path[depth].p_hdr;
658 	ex = EXT_FIRST_EXTENT(eh);
659 
660 	ext_debug("Displaying leaf extents for inode %lu\n", inode->i_ino);
661 
662 	for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ex++) {
663 		ext_debug("%d:[%d]%d:%llu ", le32_to_cpu(ex->ee_block),
664 			  ext4_ext_is_unwritten(ex),
665 			  ext4_ext_get_actual_len(ex), ext4_ext_pblock(ex));
666 	}
667 	ext_debug("\n");
668 }
669 
670 static void ext4_ext_show_move(struct inode *inode, struct ext4_ext_path *path,
671 			ext4_fsblk_t newblock, int level)
672 {
673 	int depth = ext_depth(inode);
674 	struct ext4_extent *ex;
675 
676 	if (depth != level) {
677 		struct ext4_extent_idx *idx;
678 		idx = path[level].p_idx;
679 		while (idx <= EXT_MAX_INDEX(path[level].p_hdr)) {
680 			ext_debug("%d: move %d:%llu in new index %llu\n", level,
681 					le32_to_cpu(idx->ei_block),
682 					ext4_idx_pblock(idx),
683 					newblock);
684 			idx++;
685 		}
686 
687 		return;
688 	}
689 
690 	ex = path[depth].p_ext;
691 	while (ex <= EXT_MAX_EXTENT(path[depth].p_hdr)) {
692 		ext_debug("move %d:%llu:[%d]%d in new leaf %llu\n",
693 				le32_to_cpu(ex->ee_block),
694 				ext4_ext_pblock(ex),
695 				ext4_ext_is_unwritten(ex),
696 				ext4_ext_get_actual_len(ex),
697 				newblock);
698 		ex++;
699 	}
700 }
701 
702 #else
703 #define ext4_ext_show_path(inode, path)
704 #define ext4_ext_show_leaf(inode, path)
705 #define ext4_ext_show_move(inode, path, newblock, level)
706 #endif
707 
708 void ext4_ext_drop_refs(struct ext4_ext_path *path)
709 {
710 	int depth, i;
711 
712 	if (!path)
713 		return;
714 	depth = path->p_depth;
715 	for (i = 0; i <= depth; i++, path++)
716 		if (path->p_bh) {
717 			brelse(path->p_bh);
718 			path->p_bh = NULL;
719 		}
720 }
721 
722 /*
723  * ext4_ext_binsearch_idx:
724  * binary search for the closest index of the given block
725  * the header must be checked before calling this
726  */
727 static void
728 ext4_ext_binsearch_idx(struct inode *inode,
729 			struct ext4_ext_path *path, ext4_lblk_t block)
730 {
731 	struct ext4_extent_header *eh = path->p_hdr;
732 	struct ext4_extent_idx *r, *l, *m;
733 
734 
735 	ext_debug("binsearch for %u(idx):  ", block);
736 
737 	l = EXT_FIRST_INDEX(eh) + 1;
738 	r = EXT_LAST_INDEX(eh);
739 	while (l <= r) {
740 		m = l + (r - l) / 2;
741 		if (block < le32_to_cpu(m->ei_block))
742 			r = m - 1;
743 		else
744 			l = m + 1;
745 		ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ei_block),
746 				m, le32_to_cpu(m->ei_block),
747 				r, le32_to_cpu(r->ei_block));
748 	}
749 
750 	path->p_idx = l - 1;
751 	ext_debug("  -> %u->%lld ", le32_to_cpu(path->p_idx->ei_block),
752 		  ext4_idx_pblock(path->p_idx));
753 
754 #ifdef CHECK_BINSEARCH
755 	{
756 		struct ext4_extent_idx *chix, *ix;
757 		int k;
758 
759 		chix = ix = EXT_FIRST_INDEX(eh);
760 		for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ix++) {
761 		  if (k != 0 &&
762 		      le32_to_cpu(ix->ei_block) <= le32_to_cpu(ix[-1].ei_block)) {
763 				printk(KERN_DEBUG "k=%d, ix=0x%p, "
764 				       "first=0x%p\n", k,
765 				       ix, EXT_FIRST_INDEX(eh));
766 				printk(KERN_DEBUG "%u <= %u\n",
767 				       le32_to_cpu(ix->ei_block),
768 				       le32_to_cpu(ix[-1].ei_block));
769 			}
770 			BUG_ON(k && le32_to_cpu(ix->ei_block)
771 					   <= le32_to_cpu(ix[-1].ei_block));
772 			if (block < le32_to_cpu(ix->ei_block))
773 				break;
774 			chix = ix;
775 		}
776 		BUG_ON(chix != path->p_idx);
777 	}
778 #endif
779 
780 }
781 
782 /*
783  * ext4_ext_binsearch:
784  * binary search for closest extent of the given block
785  * the header must be checked before calling this
786  */
787 static void
788 ext4_ext_binsearch(struct inode *inode,
789 		struct ext4_ext_path *path, ext4_lblk_t block)
790 {
791 	struct ext4_extent_header *eh = path->p_hdr;
792 	struct ext4_extent *r, *l, *m;
793 
794 	if (eh->eh_entries == 0) {
795 		/*
796 		 * this leaf is empty:
797 		 * we get such a leaf in split/add case
798 		 */
799 		return;
800 	}
801 
802 	ext_debug("binsearch for %u:  ", block);
803 
804 	l = EXT_FIRST_EXTENT(eh) + 1;
805 	r = EXT_LAST_EXTENT(eh);
806 
807 	while (l <= r) {
808 		m = l + (r - l) / 2;
809 		if (block < le32_to_cpu(m->ee_block))
810 			r = m - 1;
811 		else
812 			l = m + 1;
813 		ext_debug("%p(%u):%p(%u):%p(%u) ", l, le32_to_cpu(l->ee_block),
814 				m, le32_to_cpu(m->ee_block),
815 				r, le32_to_cpu(r->ee_block));
816 	}
817 
818 	path->p_ext = l - 1;
819 	ext_debug("  -> %d:%llu:[%d]%d ",
820 			le32_to_cpu(path->p_ext->ee_block),
821 			ext4_ext_pblock(path->p_ext),
822 			ext4_ext_is_unwritten(path->p_ext),
823 			ext4_ext_get_actual_len(path->p_ext));
824 
825 #ifdef CHECK_BINSEARCH
826 	{
827 		struct ext4_extent *chex, *ex;
828 		int k;
829 
830 		chex = ex = EXT_FIRST_EXTENT(eh);
831 		for (k = 0; k < le16_to_cpu(eh->eh_entries); k++, ex++) {
832 			BUG_ON(k && le32_to_cpu(ex->ee_block)
833 					  <= le32_to_cpu(ex[-1].ee_block));
834 			if (block < le32_to_cpu(ex->ee_block))
835 				break;
836 			chex = ex;
837 		}
838 		BUG_ON(chex != path->p_ext);
839 	}
840 #endif
841 
842 }
843 
844 int ext4_ext_tree_init(handle_t *handle, struct inode *inode)
845 {
846 	struct ext4_extent_header *eh;
847 
848 	eh = ext_inode_hdr(inode);
849 	eh->eh_depth = 0;
850 	eh->eh_entries = 0;
851 	eh->eh_magic = EXT4_EXT_MAGIC;
852 	eh->eh_max = cpu_to_le16(ext4_ext_space_root(inode, 0));
853 	ext4_mark_inode_dirty(handle, inode);
854 	return 0;
855 }
856 
857 struct ext4_ext_path *
858 ext4_find_extent(struct inode *inode, ext4_lblk_t block,
859 		 struct ext4_ext_path **orig_path, int flags)
860 {
861 	struct ext4_extent_header *eh;
862 	struct buffer_head *bh;
863 	struct ext4_ext_path *path = orig_path ? *orig_path : NULL;
864 	short int depth, i, ppos = 0;
865 	int ret;
866 
867 	eh = ext_inode_hdr(inode);
868 	depth = ext_depth(inode);
869 
870 	if (path) {
871 		ext4_ext_drop_refs(path);
872 		if (depth > path[0].p_maxdepth) {
873 			kfree(path);
874 			*orig_path = path = NULL;
875 		}
876 	}
877 	if (!path) {
878 		/* account possible depth increase */
879 		path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 2),
880 				GFP_NOFS);
881 		if (unlikely(!path))
882 			return ERR_PTR(-ENOMEM);
883 		path[0].p_maxdepth = depth + 1;
884 	}
885 	path[0].p_hdr = eh;
886 	path[0].p_bh = NULL;
887 
888 	i = depth;
889 	/* walk through the tree */
890 	while (i) {
891 		ext_debug("depth %d: num %d, max %d\n",
892 			  ppos, le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
893 
894 		ext4_ext_binsearch_idx(inode, path + ppos, block);
895 		path[ppos].p_block = ext4_idx_pblock(path[ppos].p_idx);
896 		path[ppos].p_depth = i;
897 		path[ppos].p_ext = NULL;
898 
899 		bh = read_extent_tree_block(inode, path[ppos].p_block, --i,
900 					    flags);
901 		if (unlikely(IS_ERR(bh))) {
902 			ret = PTR_ERR(bh);
903 			goto err;
904 		}
905 
906 		eh = ext_block_hdr(bh);
907 		ppos++;
908 		if (unlikely(ppos > depth)) {
909 			put_bh(bh);
910 			EXT4_ERROR_INODE(inode,
911 					 "ppos %d > depth %d", ppos, depth);
912 			ret = -EIO;
913 			goto err;
914 		}
915 		path[ppos].p_bh = bh;
916 		path[ppos].p_hdr = eh;
917 	}
918 
919 	path[ppos].p_depth = i;
920 	path[ppos].p_ext = NULL;
921 	path[ppos].p_idx = NULL;
922 
923 	/* find extent */
924 	ext4_ext_binsearch(inode, path + ppos, block);
925 	/* if not an empty leaf */
926 	if (path[ppos].p_ext)
927 		path[ppos].p_block = ext4_ext_pblock(path[ppos].p_ext);
928 
929 	ext4_ext_show_path(inode, path);
930 
931 	return path;
932 
933 err:
934 	ext4_ext_drop_refs(path);
935 	kfree(path);
936 	if (orig_path)
937 		*orig_path = NULL;
938 	return ERR_PTR(ret);
939 }
940 
941 /*
942  * ext4_ext_insert_index:
943  * insert new index [@logical;@ptr] into the block at @curp;
944  * check where to insert: before @curp or after @curp
945  */
946 static int ext4_ext_insert_index(handle_t *handle, struct inode *inode,
947 				 struct ext4_ext_path *curp,
948 				 int logical, ext4_fsblk_t ptr)
949 {
950 	struct ext4_extent_idx *ix;
951 	int len, err;
952 
953 	err = ext4_ext_get_access(handle, inode, curp);
954 	if (err)
955 		return err;
956 
957 	if (unlikely(logical == le32_to_cpu(curp->p_idx->ei_block))) {
958 		EXT4_ERROR_INODE(inode,
959 				 "logical %d == ei_block %d!",
960 				 logical, le32_to_cpu(curp->p_idx->ei_block));
961 		return -EIO;
962 	}
963 
964 	if (unlikely(le16_to_cpu(curp->p_hdr->eh_entries)
965 			     >= le16_to_cpu(curp->p_hdr->eh_max))) {
966 		EXT4_ERROR_INODE(inode,
967 				 "eh_entries %d >= eh_max %d!",
968 				 le16_to_cpu(curp->p_hdr->eh_entries),
969 				 le16_to_cpu(curp->p_hdr->eh_max));
970 		return -EIO;
971 	}
972 
973 	if (logical > le32_to_cpu(curp->p_idx->ei_block)) {
974 		/* insert after */
975 		ext_debug("insert new index %d after: %llu\n", logical, ptr);
976 		ix = curp->p_idx + 1;
977 	} else {
978 		/* insert before */
979 		ext_debug("insert new index %d before: %llu\n", logical, ptr);
980 		ix = curp->p_idx;
981 	}
982 
983 	len = EXT_LAST_INDEX(curp->p_hdr) - ix + 1;
984 	BUG_ON(len < 0);
985 	if (len > 0) {
986 		ext_debug("insert new index %d: "
987 				"move %d indices from 0x%p to 0x%p\n",
988 				logical, len, ix, ix + 1);
989 		memmove(ix + 1, ix, len * sizeof(struct ext4_extent_idx));
990 	}
991 
992 	if (unlikely(ix > EXT_MAX_INDEX(curp->p_hdr))) {
993 		EXT4_ERROR_INODE(inode, "ix > EXT_MAX_INDEX!");
994 		return -EIO;
995 	}
996 
997 	ix->ei_block = cpu_to_le32(logical);
998 	ext4_idx_store_pblock(ix, ptr);
999 	le16_add_cpu(&curp->p_hdr->eh_entries, 1);
1000 
1001 	if (unlikely(ix > EXT_LAST_INDEX(curp->p_hdr))) {
1002 		EXT4_ERROR_INODE(inode, "ix > EXT_LAST_INDEX!");
1003 		return -EIO;
1004 	}
1005 
1006 	err = ext4_ext_dirty(handle, inode, curp);
1007 	ext4_std_error(inode->i_sb, err);
1008 
1009 	return err;
1010 }
1011 
1012 /*
1013  * ext4_ext_split:
1014  * inserts new subtree into the path, using free index entry
1015  * at depth @at:
1016  * - allocates all needed blocks (new leaf and all intermediate index blocks)
1017  * - makes decision where to split
1018  * - moves remaining extents and index entries (right to the split point)
1019  *   into the newly allocated blocks
1020  * - initializes subtree
1021  */
1022 static int ext4_ext_split(handle_t *handle, struct inode *inode,
1023 			  unsigned int flags,
1024 			  struct ext4_ext_path *path,
1025 			  struct ext4_extent *newext, int at)
1026 {
1027 	struct buffer_head *bh = NULL;
1028 	int depth = ext_depth(inode);
1029 	struct ext4_extent_header *neh;
1030 	struct ext4_extent_idx *fidx;
1031 	int i = at, k, m, a;
1032 	ext4_fsblk_t newblock, oldblock;
1033 	__le32 border;
1034 	ext4_fsblk_t *ablocks = NULL; /* array of allocated blocks */
1035 	int err = 0;
1036 
1037 	/* make decision: where to split? */
1038 	/* FIXME: now decision is simplest: at current extent */
1039 
1040 	/* if current leaf will be split, then we should use
1041 	 * border from split point */
1042 	if (unlikely(path[depth].p_ext > EXT_MAX_EXTENT(path[depth].p_hdr))) {
1043 		EXT4_ERROR_INODE(inode, "p_ext > EXT_MAX_EXTENT!");
1044 		return -EIO;
1045 	}
1046 	if (path[depth].p_ext != EXT_MAX_EXTENT(path[depth].p_hdr)) {
1047 		border = path[depth].p_ext[1].ee_block;
1048 		ext_debug("leaf will be split."
1049 				" next leaf starts at %d\n",
1050 				  le32_to_cpu(border));
1051 	} else {
1052 		border = newext->ee_block;
1053 		ext_debug("leaf will be added."
1054 				" next leaf starts at %d\n",
1055 				le32_to_cpu(border));
1056 	}
1057 
1058 	/*
1059 	 * If error occurs, then we break processing
1060 	 * and mark filesystem read-only. index won't
1061 	 * be inserted and tree will be in consistent
1062 	 * state. Next mount will repair buffers too.
1063 	 */
1064 
1065 	/*
1066 	 * Get array to track all allocated blocks.
1067 	 * We need this to handle errors and free blocks
1068 	 * upon them.
1069 	 */
1070 	ablocks = kzalloc(sizeof(ext4_fsblk_t) * depth, GFP_NOFS);
1071 	if (!ablocks)
1072 		return -ENOMEM;
1073 
1074 	/* allocate all needed blocks */
1075 	ext_debug("allocate %d blocks for indexes/leaf\n", depth - at);
1076 	for (a = 0; a < depth - at; a++) {
1077 		newblock = ext4_ext_new_meta_block(handle, inode, path,
1078 						   newext, &err, flags);
1079 		if (newblock == 0)
1080 			goto cleanup;
1081 		ablocks[a] = newblock;
1082 	}
1083 
1084 	/* initialize new leaf */
1085 	newblock = ablocks[--a];
1086 	if (unlikely(newblock == 0)) {
1087 		EXT4_ERROR_INODE(inode, "newblock == 0!");
1088 		err = -EIO;
1089 		goto cleanup;
1090 	}
1091 	bh = sb_getblk(inode->i_sb, newblock);
1092 	if (unlikely(!bh)) {
1093 		err = -ENOMEM;
1094 		goto cleanup;
1095 	}
1096 	lock_buffer(bh);
1097 
1098 	err = ext4_journal_get_create_access(handle, bh);
1099 	if (err)
1100 		goto cleanup;
1101 
1102 	neh = ext_block_hdr(bh);
1103 	neh->eh_entries = 0;
1104 	neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1105 	neh->eh_magic = EXT4_EXT_MAGIC;
1106 	neh->eh_depth = 0;
1107 
1108 	/* move remainder of path[depth] to the new leaf */
1109 	if (unlikely(path[depth].p_hdr->eh_entries !=
1110 		     path[depth].p_hdr->eh_max)) {
1111 		EXT4_ERROR_INODE(inode, "eh_entries %d != eh_max %d!",
1112 				 path[depth].p_hdr->eh_entries,
1113 				 path[depth].p_hdr->eh_max);
1114 		err = -EIO;
1115 		goto cleanup;
1116 	}
1117 	/* start copy from next extent */
1118 	m = EXT_MAX_EXTENT(path[depth].p_hdr) - path[depth].p_ext++;
1119 	ext4_ext_show_move(inode, path, newblock, depth);
1120 	if (m) {
1121 		struct ext4_extent *ex;
1122 		ex = EXT_FIRST_EXTENT(neh);
1123 		memmove(ex, path[depth].p_ext, sizeof(struct ext4_extent) * m);
1124 		le16_add_cpu(&neh->eh_entries, m);
1125 	}
1126 
1127 	ext4_extent_block_csum_set(inode, neh);
1128 	set_buffer_uptodate(bh);
1129 	unlock_buffer(bh);
1130 
1131 	err = ext4_handle_dirty_metadata(handle, inode, bh);
1132 	if (err)
1133 		goto cleanup;
1134 	brelse(bh);
1135 	bh = NULL;
1136 
1137 	/* correct old leaf */
1138 	if (m) {
1139 		err = ext4_ext_get_access(handle, inode, path + depth);
1140 		if (err)
1141 			goto cleanup;
1142 		le16_add_cpu(&path[depth].p_hdr->eh_entries, -m);
1143 		err = ext4_ext_dirty(handle, inode, path + depth);
1144 		if (err)
1145 			goto cleanup;
1146 
1147 	}
1148 
1149 	/* create intermediate indexes */
1150 	k = depth - at - 1;
1151 	if (unlikely(k < 0)) {
1152 		EXT4_ERROR_INODE(inode, "k %d < 0!", k);
1153 		err = -EIO;
1154 		goto cleanup;
1155 	}
1156 	if (k)
1157 		ext_debug("create %d intermediate indices\n", k);
1158 	/* insert new index into current index block */
1159 	/* current depth stored in i var */
1160 	i = depth - 1;
1161 	while (k--) {
1162 		oldblock = newblock;
1163 		newblock = ablocks[--a];
1164 		bh = sb_getblk(inode->i_sb, newblock);
1165 		if (unlikely(!bh)) {
1166 			err = -ENOMEM;
1167 			goto cleanup;
1168 		}
1169 		lock_buffer(bh);
1170 
1171 		err = ext4_journal_get_create_access(handle, bh);
1172 		if (err)
1173 			goto cleanup;
1174 
1175 		neh = ext_block_hdr(bh);
1176 		neh->eh_entries = cpu_to_le16(1);
1177 		neh->eh_magic = EXT4_EXT_MAGIC;
1178 		neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1179 		neh->eh_depth = cpu_to_le16(depth - i);
1180 		fidx = EXT_FIRST_INDEX(neh);
1181 		fidx->ei_block = border;
1182 		ext4_idx_store_pblock(fidx, oldblock);
1183 
1184 		ext_debug("int.index at %d (block %llu): %u -> %llu\n",
1185 				i, newblock, le32_to_cpu(border), oldblock);
1186 
1187 		/* move remainder of path[i] to the new index block */
1188 		if (unlikely(EXT_MAX_INDEX(path[i].p_hdr) !=
1189 					EXT_LAST_INDEX(path[i].p_hdr))) {
1190 			EXT4_ERROR_INODE(inode,
1191 					 "EXT_MAX_INDEX != EXT_LAST_INDEX ee_block %d!",
1192 					 le32_to_cpu(path[i].p_ext->ee_block));
1193 			err = -EIO;
1194 			goto cleanup;
1195 		}
1196 		/* start copy indexes */
1197 		m = EXT_MAX_INDEX(path[i].p_hdr) - path[i].p_idx++;
1198 		ext_debug("cur 0x%p, last 0x%p\n", path[i].p_idx,
1199 				EXT_MAX_INDEX(path[i].p_hdr));
1200 		ext4_ext_show_move(inode, path, newblock, i);
1201 		if (m) {
1202 			memmove(++fidx, path[i].p_idx,
1203 				sizeof(struct ext4_extent_idx) * m);
1204 			le16_add_cpu(&neh->eh_entries, m);
1205 		}
1206 		ext4_extent_block_csum_set(inode, neh);
1207 		set_buffer_uptodate(bh);
1208 		unlock_buffer(bh);
1209 
1210 		err = ext4_handle_dirty_metadata(handle, inode, bh);
1211 		if (err)
1212 			goto cleanup;
1213 		brelse(bh);
1214 		bh = NULL;
1215 
1216 		/* correct old index */
1217 		if (m) {
1218 			err = ext4_ext_get_access(handle, inode, path + i);
1219 			if (err)
1220 				goto cleanup;
1221 			le16_add_cpu(&path[i].p_hdr->eh_entries, -m);
1222 			err = ext4_ext_dirty(handle, inode, path + i);
1223 			if (err)
1224 				goto cleanup;
1225 		}
1226 
1227 		i--;
1228 	}
1229 
1230 	/* insert new index */
1231 	err = ext4_ext_insert_index(handle, inode, path + at,
1232 				    le32_to_cpu(border), newblock);
1233 
1234 cleanup:
1235 	if (bh) {
1236 		if (buffer_locked(bh))
1237 			unlock_buffer(bh);
1238 		brelse(bh);
1239 	}
1240 
1241 	if (err) {
1242 		/* free all allocated blocks in error case */
1243 		for (i = 0; i < depth; i++) {
1244 			if (!ablocks[i])
1245 				continue;
1246 			ext4_free_blocks(handle, inode, NULL, ablocks[i], 1,
1247 					 EXT4_FREE_BLOCKS_METADATA);
1248 		}
1249 	}
1250 	kfree(ablocks);
1251 
1252 	return err;
1253 }
1254 
1255 /*
1256  * ext4_ext_grow_indepth:
1257  * implements tree growing procedure:
1258  * - allocates new block
1259  * - moves top-level data (index block or leaf) into the new block
1260  * - initializes new top-level, creating index that points to the
1261  *   just created block
1262  */
1263 static int ext4_ext_grow_indepth(handle_t *handle, struct inode *inode,
1264 				 unsigned int flags)
1265 {
1266 	struct ext4_extent_header *neh;
1267 	struct buffer_head *bh;
1268 	ext4_fsblk_t newblock, goal = 0;
1269 	struct ext4_super_block *es = EXT4_SB(inode->i_sb)->s_es;
1270 	int err = 0;
1271 
1272 	/* Try to prepend new index to old one */
1273 	if (ext_depth(inode))
1274 		goal = ext4_idx_pblock(EXT_FIRST_INDEX(ext_inode_hdr(inode)));
1275 	if (goal > le32_to_cpu(es->s_first_data_block)) {
1276 		flags |= EXT4_MB_HINT_TRY_GOAL;
1277 		goal--;
1278 	} else
1279 		goal = ext4_inode_to_goal_block(inode);
1280 	newblock = ext4_new_meta_blocks(handle, inode, goal, flags,
1281 					NULL, &err);
1282 	if (newblock == 0)
1283 		return err;
1284 
1285 	bh = sb_getblk(inode->i_sb, newblock);
1286 	if (unlikely(!bh))
1287 		return -ENOMEM;
1288 	lock_buffer(bh);
1289 
1290 	err = ext4_journal_get_create_access(handle, bh);
1291 	if (err) {
1292 		unlock_buffer(bh);
1293 		goto out;
1294 	}
1295 
1296 	/* move top-level index/leaf into new block */
1297 	memmove(bh->b_data, EXT4_I(inode)->i_data,
1298 		sizeof(EXT4_I(inode)->i_data));
1299 
1300 	/* set size of new block */
1301 	neh = ext_block_hdr(bh);
1302 	/* old root could have indexes or leaves
1303 	 * so calculate e_max right way */
1304 	if (ext_depth(inode))
1305 		neh->eh_max = cpu_to_le16(ext4_ext_space_block_idx(inode, 0));
1306 	else
1307 		neh->eh_max = cpu_to_le16(ext4_ext_space_block(inode, 0));
1308 	neh->eh_magic = EXT4_EXT_MAGIC;
1309 	ext4_extent_block_csum_set(inode, neh);
1310 	set_buffer_uptodate(bh);
1311 	unlock_buffer(bh);
1312 
1313 	err = ext4_handle_dirty_metadata(handle, inode, bh);
1314 	if (err)
1315 		goto out;
1316 
1317 	/* Update top-level index: num,max,pointer */
1318 	neh = ext_inode_hdr(inode);
1319 	neh->eh_entries = cpu_to_le16(1);
1320 	ext4_idx_store_pblock(EXT_FIRST_INDEX(neh), newblock);
1321 	if (neh->eh_depth == 0) {
1322 		/* Root extent block becomes index block */
1323 		neh->eh_max = cpu_to_le16(ext4_ext_space_root_idx(inode, 0));
1324 		EXT_FIRST_INDEX(neh)->ei_block =
1325 			EXT_FIRST_EXTENT(neh)->ee_block;
1326 	}
1327 	ext_debug("new root: num %d(%d), lblock %d, ptr %llu\n",
1328 		  le16_to_cpu(neh->eh_entries), le16_to_cpu(neh->eh_max),
1329 		  le32_to_cpu(EXT_FIRST_INDEX(neh)->ei_block),
1330 		  ext4_idx_pblock(EXT_FIRST_INDEX(neh)));
1331 
1332 	le16_add_cpu(&neh->eh_depth, 1);
1333 	ext4_mark_inode_dirty(handle, inode);
1334 out:
1335 	brelse(bh);
1336 
1337 	return err;
1338 }
1339 
1340 /*
1341  * ext4_ext_create_new_leaf:
1342  * finds empty index and adds new leaf.
1343  * if no free index is found, then it requests in-depth growing.
1344  */
1345 static int ext4_ext_create_new_leaf(handle_t *handle, struct inode *inode,
1346 				    unsigned int mb_flags,
1347 				    unsigned int gb_flags,
1348 				    struct ext4_ext_path **ppath,
1349 				    struct ext4_extent *newext)
1350 {
1351 	struct ext4_ext_path *path = *ppath;
1352 	struct ext4_ext_path *curp;
1353 	int depth, i, err = 0;
1354 
1355 repeat:
1356 	i = depth = ext_depth(inode);
1357 
1358 	/* walk up to the tree and look for free index entry */
1359 	curp = path + depth;
1360 	while (i > 0 && !EXT_HAS_FREE_INDEX(curp)) {
1361 		i--;
1362 		curp--;
1363 	}
1364 
1365 	/* we use already allocated block for index block,
1366 	 * so subsequent data blocks should be contiguous */
1367 	if (EXT_HAS_FREE_INDEX(curp)) {
1368 		/* if we found index with free entry, then use that
1369 		 * entry: create all needed subtree and add new leaf */
1370 		err = ext4_ext_split(handle, inode, mb_flags, path, newext, i);
1371 		if (err)
1372 			goto out;
1373 
1374 		/* refill path */
1375 		path = ext4_find_extent(inode,
1376 				    (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1377 				    ppath, gb_flags);
1378 		if (IS_ERR(path))
1379 			err = PTR_ERR(path);
1380 	} else {
1381 		/* tree is full, time to grow in depth */
1382 		err = ext4_ext_grow_indepth(handle, inode, mb_flags);
1383 		if (err)
1384 			goto out;
1385 
1386 		/* refill path */
1387 		path = ext4_find_extent(inode,
1388 				   (ext4_lblk_t)le32_to_cpu(newext->ee_block),
1389 				    ppath, gb_flags);
1390 		if (IS_ERR(path)) {
1391 			err = PTR_ERR(path);
1392 			goto out;
1393 		}
1394 
1395 		/*
1396 		 * only first (depth 0 -> 1) produces free space;
1397 		 * in all other cases we have to split the grown tree
1398 		 */
1399 		depth = ext_depth(inode);
1400 		if (path[depth].p_hdr->eh_entries == path[depth].p_hdr->eh_max) {
1401 			/* now we need to split */
1402 			goto repeat;
1403 		}
1404 	}
1405 
1406 out:
1407 	return err;
1408 }
1409 
1410 /*
1411  * search the closest allocated block to the left for *logical
1412  * and returns it at @logical + it's physical address at @phys
1413  * if *logical is the smallest allocated block, the function
1414  * returns 0 at @phys
1415  * return value contains 0 (success) or error code
1416  */
1417 static int ext4_ext_search_left(struct inode *inode,
1418 				struct ext4_ext_path *path,
1419 				ext4_lblk_t *logical, ext4_fsblk_t *phys)
1420 {
1421 	struct ext4_extent_idx *ix;
1422 	struct ext4_extent *ex;
1423 	int depth, ee_len;
1424 
1425 	if (unlikely(path == NULL)) {
1426 		EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1427 		return -EIO;
1428 	}
1429 	depth = path->p_depth;
1430 	*phys = 0;
1431 
1432 	if (depth == 0 && path->p_ext == NULL)
1433 		return 0;
1434 
1435 	/* usually extent in the path covers blocks smaller
1436 	 * then *logical, but it can be that extent is the
1437 	 * first one in the file */
1438 
1439 	ex = path[depth].p_ext;
1440 	ee_len = ext4_ext_get_actual_len(ex);
1441 	if (*logical < le32_to_cpu(ex->ee_block)) {
1442 		if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1443 			EXT4_ERROR_INODE(inode,
1444 					 "EXT_FIRST_EXTENT != ex *logical %d ee_block %d!",
1445 					 *logical, le32_to_cpu(ex->ee_block));
1446 			return -EIO;
1447 		}
1448 		while (--depth >= 0) {
1449 			ix = path[depth].p_idx;
1450 			if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1451 				EXT4_ERROR_INODE(inode,
1452 				  "ix (%d) != EXT_FIRST_INDEX (%d) (depth %d)!",
1453 				  ix != NULL ? le32_to_cpu(ix->ei_block) : 0,
1454 				  EXT_FIRST_INDEX(path[depth].p_hdr) != NULL ?
1455 		le32_to_cpu(EXT_FIRST_INDEX(path[depth].p_hdr)->ei_block) : 0,
1456 				  depth);
1457 				return -EIO;
1458 			}
1459 		}
1460 		return 0;
1461 	}
1462 
1463 	if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1464 		EXT4_ERROR_INODE(inode,
1465 				 "logical %d < ee_block %d + ee_len %d!",
1466 				 *logical, le32_to_cpu(ex->ee_block), ee_len);
1467 		return -EIO;
1468 	}
1469 
1470 	*logical = le32_to_cpu(ex->ee_block) + ee_len - 1;
1471 	*phys = ext4_ext_pblock(ex) + ee_len - 1;
1472 	return 0;
1473 }
1474 
1475 /*
1476  * search the closest allocated block to the right for *logical
1477  * and returns it at @logical + it's physical address at @phys
1478  * if *logical is the largest allocated block, the function
1479  * returns 0 at @phys
1480  * return value contains 0 (success) or error code
1481  */
1482 static int ext4_ext_search_right(struct inode *inode,
1483 				 struct ext4_ext_path *path,
1484 				 ext4_lblk_t *logical, ext4_fsblk_t *phys,
1485 				 struct ext4_extent **ret_ex)
1486 {
1487 	struct buffer_head *bh = NULL;
1488 	struct ext4_extent_header *eh;
1489 	struct ext4_extent_idx *ix;
1490 	struct ext4_extent *ex;
1491 	ext4_fsblk_t block;
1492 	int depth;	/* Note, NOT eh_depth; depth from top of tree */
1493 	int ee_len;
1494 
1495 	if (unlikely(path == NULL)) {
1496 		EXT4_ERROR_INODE(inode, "path == NULL *logical %d!", *logical);
1497 		return -EIO;
1498 	}
1499 	depth = path->p_depth;
1500 	*phys = 0;
1501 
1502 	if (depth == 0 && path->p_ext == NULL)
1503 		return 0;
1504 
1505 	/* usually extent in the path covers blocks smaller
1506 	 * then *logical, but it can be that extent is the
1507 	 * first one in the file */
1508 
1509 	ex = path[depth].p_ext;
1510 	ee_len = ext4_ext_get_actual_len(ex);
1511 	if (*logical < le32_to_cpu(ex->ee_block)) {
1512 		if (unlikely(EXT_FIRST_EXTENT(path[depth].p_hdr) != ex)) {
1513 			EXT4_ERROR_INODE(inode,
1514 					 "first_extent(path[%d].p_hdr) != ex",
1515 					 depth);
1516 			return -EIO;
1517 		}
1518 		while (--depth >= 0) {
1519 			ix = path[depth].p_idx;
1520 			if (unlikely(ix != EXT_FIRST_INDEX(path[depth].p_hdr))) {
1521 				EXT4_ERROR_INODE(inode,
1522 						 "ix != EXT_FIRST_INDEX *logical %d!",
1523 						 *logical);
1524 				return -EIO;
1525 			}
1526 		}
1527 		goto found_extent;
1528 	}
1529 
1530 	if (unlikely(*logical < (le32_to_cpu(ex->ee_block) + ee_len))) {
1531 		EXT4_ERROR_INODE(inode,
1532 				 "logical %d < ee_block %d + ee_len %d!",
1533 				 *logical, le32_to_cpu(ex->ee_block), ee_len);
1534 		return -EIO;
1535 	}
1536 
1537 	if (ex != EXT_LAST_EXTENT(path[depth].p_hdr)) {
1538 		/* next allocated block in this leaf */
1539 		ex++;
1540 		goto found_extent;
1541 	}
1542 
1543 	/* go up and search for index to the right */
1544 	while (--depth >= 0) {
1545 		ix = path[depth].p_idx;
1546 		if (ix != EXT_LAST_INDEX(path[depth].p_hdr))
1547 			goto got_index;
1548 	}
1549 
1550 	/* we've gone up to the root and found no index to the right */
1551 	return 0;
1552 
1553 got_index:
1554 	/* we've found index to the right, let's
1555 	 * follow it and find the closest allocated
1556 	 * block to the right */
1557 	ix++;
1558 	block = ext4_idx_pblock(ix);
1559 	while (++depth < path->p_depth) {
1560 		/* subtract from p_depth to get proper eh_depth */
1561 		bh = read_extent_tree_block(inode, block,
1562 					    path->p_depth - depth, 0);
1563 		if (IS_ERR(bh))
1564 			return PTR_ERR(bh);
1565 		eh = ext_block_hdr(bh);
1566 		ix = EXT_FIRST_INDEX(eh);
1567 		block = ext4_idx_pblock(ix);
1568 		put_bh(bh);
1569 	}
1570 
1571 	bh = read_extent_tree_block(inode, block, path->p_depth - depth, 0);
1572 	if (IS_ERR(bh))
1573 		return PTR_ERR(bh);
1574 	eh = ext_block_hdr(bh);
1575 	ex = EXT_FIRST_EXTENT(eh);
1576 found_extent:
1577 	*logical = le32_to_cpu(ex->ee_block);
1578 	*phys = ext4_ext_pblock(ex);
1579 	*ret_ex = ex;
1580 	if (bh)
1581 		put_bh(bh);
1582 	return 0;
1583 }
1584 
1585 /*
1586  * ext4_ext_next_allocated_block:
1587  * returns allocated block in subsequent extent or EXT_MAX_BLOCKS.
1588  * NOTE: it considers block number from index entry as
1589  * allocated block. Thus, index entries have to be consistent
1590  * with leaves.
1591  */
1592 ext4_lblk_t
1593 ext4_ext_next_allocated_block(struct ext4_ext_path *path)
1594 {
1595 	int depth;
1596 
1597 	BUG_ON(path == NULL);
1598 	depth = path->p_depth;
1599 
1600 	if (depth == 0 && path->p_ext == NULL)
1601 		return EXT_MAX_BLOCKS;
1602 
1603 	while (depth >= 0) {
1604 		if (depth == path->p_depth) {
1605 			/* leaf */
1606 			if (path[depth].p_ext &&
1607 				path[depth].p_ext !=
1608 					EXT_LAST_EXTENT(path[depth].p_hdr))
1609 			  return le32_to_cpu(path[depth].p_ext[1].ee_block);
1610 		} else {
1611 			/* index */
1612 			if (path[depth].p_idx !=
1613 					EXT_LAST_INDEX(path[depth].p_hdr))
1614 			  return le32_to_cpu(path[depth].p_idx[1].ei_block);
1615 		}
1616 		depth--;
1617 	}
1618 
1619 	return EXT_MAX_BLOCKS;
1620 }
1621 
1622 /*
1623  * ext4_ext_next_leaf_block:
1624  * returns first allocated block from next leaf or EXT_MAX_BLOCKS
1625  */
1626 static ext4_lblk_t ext4_ext_next_leaf_block(struct ext4_ext_path *path)
1627 {
1628 	int depth;
1629 
1630 	BUG_ON(path == NULL);
1631 	depth = path->p_depth;
1632 
1633 	/* zero-tree has no leaf blocks at all */
1634 	if (depth == 0)
1635 		return EXT_MAX_BLOCKS;
1636 
1637 	/* go to index block */
1638 	depth--;
1639 
1640 	while (depth >= 0) {
1641 		if (path[depth].p_idx !=
1642 				EXT_LAST_INDEX(path[depth].p_hdr))
1643 			return (ext4_lblk_t)
1644 				le32_to_cpu(path[depth].p_idx[1].ei_block);
1645 		depth--;
1646 	}
1647 
1648 	return EXT_MAX_BLOCKS;
1649 }
1650 
1651 /*
1652  * ext4_ext_correct_indexes:
1653  * if leaf gets modified and modified extent is first in the leaf,
1654  * then we have to correct all indexes above.
1655  * TODO: do we need to correct tree in all cases?
1656  */
1657 static int ext4_ext_correct_indexes(handle_t *handle, struct inode *inode,
1658 				struct ext4_ext_path *path)
1659 {
1660 	struct ext4_extent_header *eh;
1661 	int depth = ext_depth(inode);
1662 	struct ext4_extent *ex;
1663 	__le32 border;
1664 	int k, err = 0;
1665 
1666 	eh = path[depth].p_hdr;
1667 	ex = path[depth].p_ext;
1668 
1669 	if (unlikely(ex == NULL || eh == NULL)) {
1670 		EXT4_ERROR_INODE(inode,
1671 				 "ex %p == NULL or eh %p == NULL", ex, eh);
1672 		return -EIO;
1673 	}
1674 
1675 	if (depth == 0) {
1676 		/* there is no tree at all */
1677 		return 0;
1678 	}
1679 
1680 	if (ex != EXT_FIRST_EXTENT(eh)) {
1681 		/* we correct tree if first leaf got modified only */
1682 		return 0;
1683 	}
1684 
1685 	/*
1686 	 * TODO: we need correction if border is smaller than current one
1687 	 */
1688 	k = depth - 1;
1689 	border = path[depth].p_ext->ee_block;
1690 	err = ext4_ext_get_access(handle, inode, path + k);
1691 	if (err)
1692 		return err;
1693 	path[k].p_idx->ei_block = border;
1694 	err = ext4_ext_dirty(handle, inode, path + k);
1695 	if (err)
1696 		return err;
1697 
1698 	while (k--) {
1699 		/* change all left-side indexes */
1700 		if (path[k+1].p_idx != EXT_FIRST_INDEX(path[k+1].p_hdr))
1701 			break;
1702 		err = ext4_ext_get_access(handle, inode, path + k);
1703 		if (err)
1704 			break;
1705 		path[k].p_idx->ei_block = border;
1706 		err = ext4_ext_dirty(handle, inode, path + k);
1707 		if (err)
1708 			break;
1709 	}
1710 
1711 	return err;
1712 }
1713 
1714 int
1715 ext4_can_extents_be_merged(struct inode *inode, struct ext4_extent *ex1,
1716 				struct ext4_extent *ex2)
1717 {
1718 	unsigned short ext1_ee_len, ext2_ee_len;
1719 
1720 	/*
1721 	 * Make sure that both extents are initialized. We don't merge
1722 	 * unwritten extents so that we can be sure that end_io code has
1723 	 * the extent that was written properly split out and conversion to
1724 	 * initialized is trivial.
1725 	 */
1726 	if (ext4_ext_is_unwritten(ex1) != ext4_ext_is_unwritten(ex2))
1727 		return 0;
1728 
1729 	ext1_ee_len = ext4_ext_get_actual_len(ex1);
1730 	ext2_ee_len = ext4_ext_get_actual_len(ex2);
1731 
1732 	if (le32_to_cpu(ex1->ee_block) + ext1_ee_len !=
1733 			le32_to_cpu(ex2->ee_block))
1734 		return 0;
1735 
1736 	/*
1737 	 * To allow future support for preallocated extents to be added
1738 	 * as an RO_COMPAT feature, refuse to merge to extents if
1739 	 * this can result in the top bit of ee_len being set.
1740 	 */
1741 	if (ext1_ee_len + ext2_ee_len > EXT_INIT_MAX_LEN)
1742 		return 0;
1743 	if (ext4_ext_is_unwritten(ex1) &&
1744 	    (ext4_test_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN) ||
1745 	     atomic_read(&EXT4_I(inode)->i_unwritten) ||
1746 	     (ext1_ee_len + ext2_ee_len > EXT_UNWRITTEN_MAX_LEN)))
1747 		return 0;
1748 #ifdef AGGRESSIVE_TEST
1749 	if (ext1_ee_len >= 4)
1750 		return 0;
1751 #endif
1752 
1753 	if (ext4_ext_pblock(ex1) + ext1_ee_len == ext4_ext_pblock(ex2))
1754 		return 1;
1755 	return 0;
1756 }
1757 
1758 /*
1759  * This function tries to merge the "ex" extent to the next extent in the tree.
1760  * It always tries to merge towards right. If you want to merge towards
1761  * left, pass "ex - 1" as argument instead of "ex".
1762  * Returns 0 if the extents (ex and ex+1) were _not_ merged and returns
1763  * 1 if they got merged.
1764  */
1765 static int ext4_ext_try_to_merge_right(struct inode *inode,
1766 				 struct ext4_ext_path *path,
1767 				 struct ext4_extent *ex)
1768 {
1769 	struct ext4_extent_header *eh;
1770 	unsigned int depth, len;
1771 	int merge_done = 0, unwritten;
1772 
1773 	depth = ext_depth(inode);
1774 	BUG_ON(path[depth].p_hdr == NULL);
1775 	eh = path[depth].p_hdr;
1776 
1777 	while (ex < EXT_LAST_EXTENT(eh)) {
1778 		if (!ext4_can_extents_be_merged(inode, ex, ex + 1))
1779 			break;
1780 		/* merge with next extent! */
1781 		unwritten = ext4_ext_is_unwritten(ex);
1782 		ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1783 				+ ext4_ext_get_actual_len(ex + 1));
1784 		if (unwritten)
1785 			ext4_ext_mark_unwritten(ex);
1786 
1787 		if (ex + 1 < EXT_LAST_EXTENT(eh)) {
1788 			len = (EXT_LAST_EXTENT(eh) - ex - 1)
1789 				* sizeof(struct ext4_extent);
1790 			memmove(ex + 1, ex + 2, len);
1791 		}
1792 		le16_add_cpu(&eh->eh_entries, -1);
1793 		merge_done = 1;
1794 		WARN_ON(eh->eh_entries == 0);
1795 		if (!eh->eh_entries)
1796 			EXT4_ERROR_INODE(inode, "eh->eh_entries = 0!");
1797 	}
1798 
1799 	return merge_done;
1800 }
1801 
1802 /*
1803  * This function does a very simple check to see if we can collapse
1804  * an extent tree with a single extent tree leaf block into the inode.
1805  */
1806 static void ext4_ext_try_to_merge_up(handle_t *handle,
1807 				     struct inode *inode,
1808 				     struct ext4_ext_path *path)
1809 {
1810 	size_t s;
1811 	unsigned max_root = ext4_ext_space_root(inode, 0);
1812 	ext4_fsblk_t blk;
1813 
1814 	if ((path[0].p_depth != 1) ||
1815 	    (le16_to_cpu(path[0].p_hdr->eh_entries) != 1) ||
1816 	    (le16_to_cpu(path[1].p_hdr->eh_entries) > max_root))
1817 		return;
1818 
1819 	/*
1820 	 * We need to modify the block allocation bitmap and the block
1821 	 * group descriptor to release the extent tree block.  If we
1822 	 * can't get the journal credits, give up.
1823 	 */
1824 	if (ext4_journal_extend(handle, 2))
1825 		return;
1826 
1827 	/*
1828 	 * Copy the extent data up to the inode
1829 	 */
1830 	blk = ext4_idx_pblock(path[0].p_idx);
1831 	s = le16_to_cpu(path[1].p_hdr->eh_entries) *
1832 		sizeof(struct ext4_extent_idx);
1833 	s += sizeof(struct ext4_extent_header);
1834 
1835 	path[1].p_maxdepth = path[0].p_maxdepth;
1836 	memcpy(path[0].p_hdr, path[1].p_hdr, s);
1837 	path[0].p_depth = 0;
1838 	path[0].p_ext = EXT_FIRST_EXTENT(path[0].p_hdr) +
1839 		(path[1].p_ext - EXT_FIRST_EXTENT(path[1].p_hdr));
1840 	path[0].p_hdr->eh_max = cpu_to_le16(max_root);
1841 
1842 	brelse(path[1].p_bh);
1843 	ext4_free_blocks(handle, inode, NULL, blk, 1,
1844 			 EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
1845 }
1846 
1847 /*
1848  * This function tries to merge the @ex extent to neighbours in the tree.
1849  * return 1 if merge left else 0.
1850  */
1851 static void ext4_ext_try_to_merge(handle_t *handle,
1852 				  struct inode *inode,
1853 				  struct ext4_ext_path *path,
1854 				  struct ext4_extent *ex) {
1855 	struct ext4_extent_header *eh;
1856 	unsigned int depth;
1857 	int merge_done = 0;
1858 
1859 	depth = ext_depth(inode);
1860 	BUG_ON(path[depth].p_hdr == NULL);
1861 	eh = path[depth].p_hdr;
1862 
1863 	if (ex > EXT_FIRST_EXTENT(eh))
1864 		merge_done = ext4_ext_try_to_merge_right(inode, path, ex - 1);
1865 
1866 	if (!merge_done)
1867 		(void) ext4_ext_try_to_merge_right(inode, path, ex);
1868 
1869 	ext4_ext_try_to_merge_up(handle, inode, path);
1870 }
1871 
1872 /*
1873  * check if a portion of the "newext" extent overlaps with an
1874  * existing extent.
1875  *
1876  * If there is an overlap discovered, it updates the length of the newext
1877  * such that there will be no overlap, and then returns 1.
1878  * If there is no overlap found, it returns 0.
1879  */
1880 static unsigned int ext4_ext_check_overlap(struct ext4_sb_info *sbi,
1881 					   struct inode *inode,
1882 					   struct ext4_extent *newext,
1883 					   struct ext4_ext_path *path)
1884 {
1885 	ext4_lblk_t b1, b2;
1886 	unsigned int depth, len1;
1887 	unsigned int ret = 0;
1888 
1889 	b1 = le32_to_cpu(newext->ee_block);
1890 	len1 = ext4_ext_get_actual_len(newext);
1891 	depth = ext_depth(inode);
1892 	if (!path[depth].p_ext)
1893 		goto out;
1894 	b2 = EXT4_LBLK_CMASK(sbi, le32_to_cpu(path[depth].p_ext->ee_block));
1895 
1896 	/*
1897 	 * get the next allocated block if the extent in the path
1898 	 * is before the requested block(s)
1899 	 */
1900 	if (b2 < b1) {
1901 		b2 = ext4_ext_next_allocated_block(path);
1902 		if (b2 == EXT_MAX_BLOCKS)
1903 			goto out;
1904 		b2 = EXT4_LBLK_CMASK(sbi, b2);
1905 	}
1906 
1907 	/* check for wrap through zero on extent logical start block*/
1908 	if (b1 + len1 < b1) {
1909 		len1 = EXT_MAX_BLOCKS - b1;
1910 		newext->ee_len = cpu_to_le16(len1);
1911 		ret = 1;
1912 	}
1913 
1914 	/* check for overlap */
1915 	if (b1 + len1 > b2) {
1916 		newext->ee_len = cpu_to_le16(b2 - b1);
1917 		ret = 1;
1918 	}
1919 out:
1920 	return ret;
1921 }
1922 
1923 /*
1924  * ext4_ext_insert_extent:
1925  * tries to merge requsted extent into the existing extent or
1926  * inserts requested extent as new one into the tree,
1927  * creating new leaf in the no-space case.
1928  */
1929 int ext4_ext_insert_extent(handle_t *handle, struct inode *inode,
1930 				struct ext4_ext_path **ppath,
1931 				struct ext4_extent *newext, int gb_flags)
1932 {
1933 	struct ext4_ext_path *path = *ppath;
1934 	struct ext4_extent_header *eh;
1935 	struct ext4_extent *ex, *fex;
1936 	struct ext4_extent *nearex; /* nearest extent */
1937 	struct ext4_ext_path *npath = NULL;
1938 	int depth, len, err;
1939 	ext4_lblk_t next;
1940 	int mb_flags = 0, unwritten;
1941 
1942 	if (gb_flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1943 		mb_flags |= EXT4_MB_DELALLOC_RESERVED;
1944 	if (unlikely(ext4_ext_get_actual_len(newext) == 0)) {
1945 		EXT4_ERROR_INODE(inode, "ext4_ext_get_actual_len(newext) == 0");
1946 		return -EIO;
1947 	}
1948 	depth = ext_depth(inode);
1949 	ex = path[depth].p_ext;
1950 	eh = path[depth].p_hdr;
1951 	if (unlikely(path[depth].p_hdr == NULL)) {
1952 		EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
1953 		return -EIO;
1954 	}
1955 
1956 	/* try to insert block into found extent and return */
1957 	if (ex && !(gb_flags & EXT4_GET_BLOCKS_PRE_IO)) {
1958 
1959 		/*
1960 		 * Try to see whether we should rather test the extent on
1961 		 * right from ex, or from the left of ex. This is because
1962 		 * ext4_find_extent() can return either extent on the
1963 		 * left, or on the right from the searched position. This
1964 		 * will make merging more effective.
1965 		 */
1966 		if (ex < EXT_LAST_EXTENT(eh) &&
1967 		    (le32_to_cpu(ex->ee_block) +
1968 		    ext4_ext_get_actual_len(ex) <
1969 		    le32_to_cpu(newext->ee_block))) {
1970 			ex += 1;
1971 			goto prepend;
1972 		} else if ((ex > EXT_FIRST_EXTENT(eh)) &&
1973 			   (le32_to_cpu(newext->ee_block) +
1974 			   ext4_ext_get_actual_len(newext) <
1975 			   le32_to_cpu(ex->ee_block)))
1976 			ex -= 1;
1977 
1978 		/* Try to append newex to the ex */
1979 		if (ext4_can_extents_be_merged(inode, ex, newext)) {
1980 			ext_debug("append [%d]%d block to %u:[%d]%d"
1981 				  "(from %llu)\n",
1982 				  ext4_ext_is_unwritten(newext),
1983 				  ext4_ext_get_actual_len(newext),
1984 				  le32_to_cpu(ex->ee_block),
1985 				  ext4_ext_is_unwritten(ex),
1986 				  ext4_ext_get_actual_len(ex),
1987 				  ext4_ext_pblock(ex));
1988 			err = ext4_ext_get_access(handle, inode,
1989 						  path + depth);
1990 			if (err)
1991 				return err;
1992 			unwritten = ext4_ext_is_unwritten(ex);
1993 			ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
1994 					+ ext4_ext_get_actual_len(newext));
1995 			if (unwritten)
1996 				ext4_ext_mark_unwritten(ex);
1997 			eh = path[depth].p_hdr;
1998 			nearex = ex;
1999 			goto merge;
2000 		}
2001 
2002 prepend:
2003 		/* Try to prepend newex to the ex */
2004 		if (ext4_can_extents_be_merged(inode, newext, ex)) {
2005 			ext_debug("prepend %u[%d]%d block to %u:[%d]%d"
2006 				  "(from %llu)\n",
2007 				  le32_to_cpu(newext->ee_block),
2008 				  ext4_ext_is_unwritten(newext),
2009 				  ext4_ext_get_actual_len(newext),
2010 				  le32_to_cpu(ex->ee_block),
2011 				  ext4_ext_is_unwritten(ex),
2012 				  ext4_ext_get_actual_len(ex),
2013 				  ext4_ext_pblock(ex));
2014 			err = ext4_ext_get_access(handle, inode,
2015 						  path + depth);
2016 			if (err)
2017 				return err;
2018 
2019 			unwritten = ext4_ext_is_unwritten(ex);
2020 			ex->ee_block = newext->ee_block;
2021 			ext4_ext_store_pblock(ex, ext4_ext_pblock(newext));
2022 			ex->ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex)
2023 					+ ext4_ext_get_actual_len(newext));
2024 			if (unwritten)
2025 				ext4_ext_mark_unwritten(ex);
2026 			eh = path[depth].p_hdr;
2027 			nearex = ex;
2028 			goto merge;
2029 		}
2030 	}
2031 
2032 	depth = ext_depth(inode);
2033 	eh = path[depth].p_hdr;
2034 	if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max))
2035 		goto has_space;
2036 
2037 	/* probably next leaf has space for us? */
2038 	fex = EXT_LAST_EXTENT(eh);
2039 	next = EXT_MAX_BLOCKS;
2040 	if (le32_to_cpu(newext->ee_block) > le32_to_cpu(fex->ee_block))
2041 		next = ext4_ext_next_leaf_block(path);
2042 	if (next != EXT_MAX_BLOCKS) {
2043 		ext_debug("next leaf block - %u\n", next);
2044 		BUG_ON(npath != NULL);
2045 		npath = ext4_find_extent(inode, next, NULL, 0);
2046 		if (IS_ERR(npath))
2047 			return PTR_ERR(npath);
2048 		BUG_ON(npath->p_depth != path->p_depth);
2049 		eh = npath[depth].p_hdr;
2050 		if (le16_to_cpu(eh->eh_entries) < le16_to_cpu(eh->eh_max)) {
2051 			ext_debug("next leaf isn't full(%d)\n",
2052 				  le16_to_cpu(eh->eh_entries));
2053 			path = npath;
2054 			goto has_space;
2055 		}
2056 		ext_debug("next leaf has no free space(%d,%d)\n",
2057 			  le16_to_cpu(eh->eh_entries), le16_to_cpu(eh->eh_max));
2058 	}
2059 
2060 	/*
2061 	 * There is no free space in the found leaf.
2062 	 * We're gonna add a new leaf in the tree.
2063 	 */
2064 	if (gb_flags & EXT4_GET_BLOCKS_METADATA_NOFAIL)
2065 		mb_flags |= EXT4_MB_USE_RESERVED;
2066 	err = ext4_ext_create_new_leaf(handle, inode, mb_flags, gb_flags,
2067 				       ppath, newext);
2068 	if (err)
2069 		goto cleanup;
2070 	depth = ext_depth(inode);
2071 	eh = path[depth].p_hdr;
2072 
2073 has_space:
2074 	nearex = path[depth].p_ext;
2075 
2076 	err = ext4_ext_get_access(handle, inode, path + depth);
2077 	if (err)
2078 		goto cleanup;
2079 
2080 	if (!nearex) {
2081 		/* there is no extent in this leaf, create first one */
2082 		ext_debug("first extent in the leaf: %u:%llu:[%d]%d\n",
2083 				le32_to_cpu(newext->ee_block),
2084 				ext4_ext_pblock(newext),
2085 				ext4_ext_is_unwritten(newext),
2086 				ext4_ext_get_actual_len(newext));
2087 		nearex = EXT_FIRST_EXTENT(eh);
2088 	} else {
2089 		if (le32_to_cpu(newext->ee_block)
2090 			   > le32_to_cpu(nearex->ee_block)) {
2091 			/* Insert after */
2092 			ext_debug("insert %u:%llu:[%d]%d before: "
2093 					"nearest %p\n",
2094 					le32_to_cpu(newext->ee_block),
2095 					ext4_ext_pblock(newext),
2096 					ext4_ext_is_unwritten(newext),
2097 					ext4_ext_get_actual_len(newext),
2098 					nearex);
2099 			nearex++;
2100 		} else {
2101 			/* Insert before */
2102 			BUG_ON(newext->ee_block == nearex->ee_block);
2103 			ext_debug("insert %u:%llu:[%d]%d after: "
2104 					"nearest %p\n",
2105 					le32_to_cpu(newext->ee_block),
2106 					ext4_ext_pblock(newext),
2107 					ext4_ext_is_unwritten(newext),
2108 					ext4_ext_get_actual_len(newext),
2109 					nearex);
2110 		}
2111 		len = EXT_LAST_EXTENT(eh) - nearex + 1;
2112 		if (len > 0) {
2113 			ext_debug("insert %u:%llu:[%d]%d: "
2114 					"move %d extents from 0x%p to 0x%p\n",
2115 					le32_to_cpu(newext->ee_block),
2116 					ext4_ext_pblock(newext),
2117 					ext4_ext_is_unwritten(newext),
2118 					ext4_ext_get_actual_len(newext),
2119 					len, nearex, nearex + 1);
2120 			memmove(nearex + 1, nearex,
2121 				len * sizeof(struct ext4_extent));
2122 		}
2123 	}
2124 
2125 	le16_add_cpu(&eh->eh_entries, 1);
2126 	path[depth].p_ext = nearex;
2127 	nearex->ee_block = newext->ee_block;
2128 	ext4_ext_store_pblock(nearex, ext4_ext_pblock(newext));
2129 	nearex->ee_len = newext->ee_len;
2130 
2131 merge:
2132 	/* try to merge extents */
2133 	if (!(gb_flags & EXT4_GET_BLOCKS_PRE_IO))
2134 		ext4_ext_try_to_merge(handle, inode, path, nearex);
2135 
2136 
2137 	/* time to correct all indexes above */
2138 	err = ext4_ext_correct_indexes(handle, inode, path);
2139 	if (err)
2140 		goto cleanup;
2141 
2142 	err = ext4_ext_dirty(handle, inode, path + path->p_depth);
2143 
2144 cleanup:
2145 	ext4_ext_drop_refs(npath);
2146 	kfree(npath);
2147 	return err;
2148 }
2149 
2150 static int ext4_fill_fiemap_extents(struct inode *inode,
2151 				    ext4_lblk_t block, ext4_lblk_t num,
2152 				    struct fiemap_extent_info *fieinfo)
2153 {
2154 	struct ext4_ext_path *path = NULL;
2155 	struct ext4_extent *ex;
2156 	struct extent_status es;
2157 	ext4_lblk_t next, next_del, start = 0, end = 0;
2158 	ext4_lblk_t last = block + num;
2159 	int exists, depth = 0, err = 0;
2160 	unsigned int flags = 0;
2161 	unsigned char blksize_bits = inode->i_sb->s_blocksize_bits;
2162 
2163 	while (block < last && block != EXT_MAX_BLOCKS) {
2164 		num = last - block;
2165 		/* find extent for this block */
2166 		down_read(&EXT4_I(inode)->i_data_sem);
2167 
2168 		path = ext4_find_extent(inode, block, &path, 0);
2169 		if (IS_ERR(path)) {
2170 			up_read(&EXT4_I(inode)->i_data_sem);
2171 			err = PTR_ERR(path);
2172 			path = NULL;
2173 			break;
2174 		}
2175 
2176 		depth = ext_depth(inode);
2177 		if (unlikely(path[depth].p_hdr == NULL)) {
2178 			up_read(&EXT4_I(inode)->i_data_sem);
2179 			EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2180 			err = -EIO;
2181 			break;
2182 		}
2183 		ex = path[depth].p_ext;
2184 		next = ext4_ext_next_allocated_block(path);
2185 
2186 		flags = 0;
2187 		exists = 0;
2188 		if (!ex) {
2189 			/* there is no extent yet, so try to allocate
2190 			 * all requested space */
2191 			start = block;
2192 			end = block + num;
2193 		} else if (le32_to_cpu(ex->ee_block) > block) {
2194 			/* need to allocate space before found extent */
2195 			start = block;
2196 			end = le32_to_cpu(ex->ee_block);
2197 			if (block + num < end)
2198 				end = block + num;
2199 		} else if (block >= le32_to_cpu(ex->ee_block)
2200 					+ ext4_ext_get_actual_len(ex)) {
2201 			/* need to allocate space after found extent */
2202 			start = block;
2203 			end = block + num;
2204 			if (end >= next)
2205 				end = next;
2206 		} else if (block >= le32_to_cpu(ex->ee_block)) {
2207 			/*
2208 			 * some part of requested space is covered
2209 			 * by found extent
2210 			 */
2211 			start = block;
2212 			end = le32_to_cpu(ex->ee_block)
2213 				+ ext4_ext_get_actual_len(ex);
2214 			if (block + num < end)
2215 				end = block + num;
2216 			exists = 1;
2217 		} else {
2218 			BUG();
2219 		}
2220 		BUG_ON(end <= start);
2221 
2222 		if (!exists) {
2223 			es.es_lblk = start;
2224 			es.es_len = end - start;
2225 			es.es_pblk = 0;
2226 		} else {
2227 			es.es_lblk = le32_to_cpu(ex->ee_block);
2228 			es.es_len = ext4_ext_get_actual_len(ex);
2229 			es.es_pblk = ext4_ext_pblock(ex);
2230 			if (ext4_ext_is_unwritten(ex))
2231 				flags |= FIEMAP_EXTENT_UNWRITTEN;
2232 		}
2233 
2234 		/*
2235 		 * Find delayed extent and update es accordingly. We call
2236 		 * it even in !exists case to find out whether es is the
2237 		 * last existing extent or not.
2238 		 */
2239 		next_del = ext4_find_delayed_extent(inode, &es);
2240 		if (!exists && next_del) {
2241 			exists = 1;
2242 			flags |= (FIEMAP_EXTENT_DELALLOC |
2243 				  FIEMAP_EXTENT_UNKNOWN);
2244 		}
2245 		up_read(&EXT4_I(inode)->i_data_sem);
2246 
2247 		if (unlikely(es.es_len == 0)) {
2248 			EXT4_ERROR_INODE(inode, "es.es_len == 0");
2249 			err = -EIO;
2250 			break;
2251 		}
2252 
2253 		/*
2254 		 * This is possible iff next == next_del == EXT_MAX_BLOCKS.
2255 		 * we need to check next == EXT_MAX_BLOCKS because it is
2256 		 * possible that an extent is with unwritten and delayed
2257 		 * status due to when an extent is delayed allocated and
2258 		 * is allocated by fallocate status tree will track both of
2259 		 * them in a extent.
2260 		 *
2261 		 * So we could return a unwritten and delayed extent, and
2262 		 * its block is equal to 'next'.
2263 		 */
2264 		if (next == next_del && next == EXT_MAX_BLOCKS) {
2265 			flags |= FIEMAP_EXTENT_LAST;
2266 			if (unlikely(next_del != EXT_MAX_BLOCKS ||
2267 				     next != EXT_MAX_BLOCKS)) {
2268 				EXT4_ERROR_INODE(inode,
2269 						 "next extent == %u, next "
2270 						 "delalloc extent = %u",
2271 						 next, next_del);
2272 				err = -EIO;
2273 				break;
2274 			}
2275 		}
2276 
2277 		if (exists) {
2278 			err = fiemap_fill_next_extent(fieinfo,
2279 				(__u64)es.es_lblk << blksize_bits,
2280 				(__u64)es.es_pblk << blksize_bits,
2281 				(__u64)es.es_len << blksize_bits,
2282 				flags);
2283 			if (err < 0)
2284 				break;
2285 			if (err == 1) {
2286 				err = 0;
2287 				break;
2288 			}
2289 		}
2290 
2291 		block = es.es_lblk + es.es_len;
2292 	}
2293 
2294 	ext4_ext_drop_refs(path);
2295 	kfree(path);
2296 	return err;
2297 }
2298 
2299 /*
2300  * ext4_ext_put_gap_in_cache:
2301  * calculate boundaries of the gap that the requested block fits into
2302  * and cache this gap
2303  */
2304 static void
2305 ext4_ext_put_gap_in_cache(struct inode *inode, struct ext4_ext_path *path,
2306 				ext4_lblk_t block)
2307 {
2308 	int depth = ext_depth(inode);
2309 	unsigned long len = 0;
2310 	ext4_lblk_t lblock = 0;
2311 	struct ext4_extent *ex;
2312 
2313 	ex = path[depth].p_ext;
2314 	if (ex == NULL) {
2315 		/*
2316 		 * there is no extent yet, so gap is [0;-] and we
2317 		 * don't cache it
2318 		 */
2319 		ext_debug("cache gap(whole file):");
2320 	} else if (block < le32_to_cpu(ex->ee_block)) {
2321 		lblock = block;
2322 		len = le32_to_cpu(ex->ee_block) - block;
2323 		ext_debug("cache gap(before): %u [%u:%u]",
2324 				block,
2325 				le32_to_cpu(ex->ee_block),
2326 				 ext4_ext_get_actual_len(ex));
2327 		if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2328 			ext4_es_insert_extent(inode, lblock, len, ~0,
2329 					      EXTENT_STATUS_HOLE);
2330 	} else if (block >= le32_to_cpu(ex->ee_block)
2331 			+ ext4_ext_get_actual_len(ex)) {
2332 		ext4_lblk_t next;
2333 		lblock = le32_to_cpu(ex->ee_block)
2334 			+ ext4_ext_get_actual_len(ex);
2335 
2336 		next = ext4_ext_next_allocated_block(path);
2337 		ext_debug("cache gap(after): [%u:%u] %u",
2338 				le32_to_cpu(ex->ee_block),
2339 				ext4_ext_get_actual_len(ex),
2340 				block);
2341 		BUG_ON(next == lblock);
2342 		len = next - lblock;
2343 		if (!ext4_find_delalloc_range(inode, lblock, lblock + len - 1))
2344 			ext4_es_insert_extent(inode, lblock, len, ~0,
2345 					      EXTENT_STATUS_HOLE);
2346 	} else {
2347 		BUG();
2348 	}
2349 
2350 	ext_debug(" -> %u:%lu\n", lblock, len);
2351 }
2352 
2353 /*
2354  * ext4_ext_rm_idx:
2355  * removes index from the index block.
2356  */
2357 static int ext4_ext_rm_idx(handle_t *handle, struct inode *inode,
2358 			struct ext4_ext_path *path, int depth)
2359 {
2360 	int err;
2361 	ext4_fsblk_t leaf;
2362 
2363 	/* free index block */
2364 	depth--;
2365 	path = path + depth;
2366 	leaf = ext4_idx_pblock(path->p_idx);
2367 	if (unlikely(path->p_hdr->eh_entries == 0)) {
2368 		EXT4_ERROR_INODE(inode, "path->p_hdr->eh_entries == 0");
2369 		return -EIO;
2370 	}
2371 	err = ext4_ext_get_access(handle, inode, path);
2372 	if (err)
2373 		return err;
2374 
2375 	if (path->p_idx != EXT_LAST_INDEX(path->p_hdr)) {
2376 		int len = EXT_LAST_INDEX(path->p_hdr) - path->p_idx;
2377 		len *= sizeof(struct ext4_extent_idx);
2378 		memmove(path->p_idx, path->p_idx + 1, len);
2379 	}
2380 
2381 	le16_add_cpu(&path->p_hdr->eh_entries, -1);
2382 	err = ext4_ext_dirty(handle, inode, path);
2383 	if (err)
2384 		return err;
2385 	ext_debug("index is empty, remove it, free block %llu\n", leaf);
2386 	trace_ext4_ext_rm_idx(inode, leaf);
2387 
2388 	ext4_free_blocks(handle, inode, NULL, leaf, 1,
2389 			 EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET);
2390 
2391 	while (--depth >= 0) {
2392 		if (path->p_idx != EXT_FIRST_INDEX(path->p_hdr))
2393 			break;
2394 		path--;
2395 		err = ext4_ext_get_access(handle, inode, path);
2396 		if (err)
2397 			break;
2398 		path->p_idx->ei_block = (path+1)->p_idx->ei_block;
2399 		err = ext4_ext_dirty(handle, inode, path);
2400 		if (err)
2401 			break;
2402 	}
2403 	return err;
2404 }
2405 
2406 /*
2407  * ext4_ext_calc_credits_for_single_extent:
2408  * This routine returns max. credits that needed to insert an extent
2409  * to the extent tree.
2410  * When pass the actual path, the caller should calculate credits
2411  * under i_data_sem.
2412  */
2413 int ext4_ext_calc_credits_for_single_extent(struct inode *inode, int nrblocks,
2414 						struct ext4_ext_path *path)
2415 {
2416 	if (path) {
2417 		int depth = ext_depth(inode);
2418 		int ret = 0;
2419 
2420 		/* probably there is space in leaf? */
2421 		if (le16_to_cpu(path[depth].p_hdr->eh_entries)
2422 				< le16_to_cpu(path[depth].p_hdr->eh_max)) {
2423 
2424 			/*
2425 			 *  There are some space in the leaf tree, no
2426 			 *  need to account for leaf block credit
2427 			 *
2428 			 *  bitmaps and block group descriptor blocks
2429 			 *  and other metadata blocks still need to be
2430 			 *  accounted.
2431 			 */
2432 			/* 1 bitmap, 1 block group descriptor */
2433 			ret = 2 + EXT4_META_TRANS_BLOCKS(inode->i_sb);
2434 			return ret;
2435 		}
2436 	}
2437 
2438 	return ext4_chunk_trans_blocks(inode, nrblocks);
2439 }
2440 
2441 /*
2442  * How many index/leaf blocks need to change/allocate to add @extents extents?
2443  *
2444  * If we add a single extent, then in the worse case, each tree level
2445  * index/leaf need to be changed in case of the tree split.
2446  *
2447  * If more extents are inserted, they could cause the whole tree split more
2448  * than once, but this is really rare.
2449  */
2450 int ext4_ext_index_trans_blocks(struct inode *inode, int extents)
2451 {
2452 	int index;
2453 	int depth;
2454 
2455 	/* If we are converting the inline data, only one is needed here. */
2456 	if (ext4_has_inline_data(inode))
2457 		return 1;
2458 
2459 	depth = ext_depth(inode);
2460 
2461 	if (extents <= 1)
2462 		index = depth * 2;
2463 	else
2464 		index = depth * 3;
2465 
2466 	return index;
2467 }
2468 
2469 static inline int get_default_free_blocks_flags(struct inode *inode)
2470 {
2471 	if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
2472 		return EXT4_FREE_BLOCKS_METADATA | EXT4_FREE_BLOCKS_FORGET;
2473 	else if (ext4_should_journal_data(inode))
2474 		return EXT4_FREE_BLOCKS_FORGET;
2475 	return 0;
2476 }
2477 
2478 static int ext4_remove_blocks(handle_t *handle, struct inode *inode,
2479 			      struct ext4_extent *ex,
2480 			      long long *partial_cluster,
2481 			      ext4_lblk_t from, ext4_lblk_t to)
2482 {
2483 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2484 	unsigned short ee_len =  ext4_ext_get_actual_len(ex);
2485 	ext4_fsblk_t pblk;
2486 	int flags = get_default_free_blocks_flags(inode);
2487 
2488 	/*
2489 	 * For bigalloc file systems, we never free a partial cluster
2490 	 * at the beginning of the extent.  Instead, we make a note
2491 	 * that we tried freeing the cluster, and check to see if we
2492 	 * need to free it on a subsequent call to ext4_remove_blocks,
2493 	 * or at the end of the ext4_truncate() operation.
2494 	 */
2495 	flags |= EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER;
2496 
2497 	trace_ext4_remove_blocks(inode, ex, from, to, *partial_cluster);
2498 	/*
2499 	 * If we have a partial cluster, and it's different from the
2500 	 * cluster of the last block, we need to explicitly free the
2501 	 * partial cluster here.
2502 	 */
2503 	pblk = ext4_ext_pblock(ex) + ee_len - 1;
2504 	if ((*partial_cluster > 0) &&
2505 	    (EXT4_B2C(sbi, pblk) != *partial_cluster)) {
2506 		ext4_free_blocks(handle, inode, NULL,
2507 				 EXT4_C2B(sbi, *partial_cluster),
2508 				 sbi->s_cluster_ratio, flags);
2509 		*partial_cluster = 0;
2510 	}
2511 
2512 #ifdef EXTENTS_STATS
2513 	{
2514 		struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2515 		spin_lock(&sbi->s_ext_stats_lock);
2516 		sbi->s_ext_blocks += ee_len;
2517 		sbi->s_ext_extents++;
2518 		if (ee_len < sbi->s_ext_min)
2519 			sbi->s_ext_min = ee_len;
2520 		if (ee_len > sbi->s_ext_max)
2521 			sbi->s_ext_max = ee_len;
2522 		if (ext_depth(inode) > sbi->s_depth_max)
2523 			sbi->s_depth_max = ext_depth(inode);
2524 		spin_unlock(&sbi->s_ext_stats_lock);
2525 	}
2526 #endif
2527 	if (from >= le32_to_cpu(ex->ee_block)
2528 	    && to == le32_to_cpu(ex->ee_block) + ee_len - 1) {
2529 		/* tail removal */
2530 		ext4_lblk_t num;
2531 		unsigned int unaligned;
2532 
2533 		num = le32_to_cpu(ex->ee_block) + ee_len - from;
2534 		pblk = ext4_ext_pblock(ex) + ee_len - num;
2535 		/*
2536 		 * Usually we want to free partial cluster at the end of the
2537 		 * extent, except for the situation when the cluster is still
2538 		 * used by any other extent (partial_cluster is negative).
2539 		 */
2540 		if (*partial_cluster < 0 &&
2541 		    -(*partial_cluster) == EXT4_B2C(sbi, pblk + num - 1))
2542 			flags |= EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER;
2543 
2544 		ext_debug("free last %u blocks starting %llu partial %lld\n",
2545 			  num, pblk, *partial_cluster);
2546 		ext4_free_blocks(handle, inode, NULL, pblk, num, flags);
2547 		/*
2548 		 * If the block range to be freed didn't start at the
2549 		 * beginning of a cluster, and we removed the entire
2550 		 * extent and the cluster is not used by any other extent,
2551 		 * save the partial cluster here, since we might need to
2552 		 * delete if we determine that the truncate operation has
2553 		 * removed all of the blocks in the cluster.
2554 		 *
2555 		 * On the other hand, if we did not manage to free the whole
2556 		 * extent, we have to mark the cluster as used (store negative
2557 		 * cluster number in partial_cluster).
2558 		 */
2559 		unaligned = EXT4_PBLK_COFF(sbi, pblk);
2560 		if (unaligned && (ee_len == num) &&
2561 		    (*partial_cluster != -((long long)EXT4_B2C(sbi, pblk))))
2562 			*partial_cluster = EXT4_B2C(sbi, pblk);
2563 		else if (unaligned)
2564 			*partial_cluster = -((long long)EXT4_B2C(sbi, pblk));
2565 		else if (*partial_cluster > 0)
2566 			*partial_cluster = 0;
2567 	} else
2568 		ext4_error(sbi->s_sb, "strange request: removal(2) "
2569 			   "%u-%u from %u:%u\n",
2570 			   from, to, le32_to_cpu(ex->ee_block), ee_len);
2571 	return 0;
2572 }
2573 
2574 
2575 /*
2576  * ext4_ext_rm_leaf() Removes the extents associated with the
2577  * blocks appearing between "start" and "end", and splits the extents
2578  * if "start" and "end" appear in the same extent
2579  *
2580  * @handle: The journal handle
2581  * @inode:  The files inode
2582  * @path:   The path to the leaf
2583  * @partial_cluster: The cluster which we'll have to free if all extents
2584  *                   has been released from it. It gets negative in case
2585  *                   that the cluster is still used.
2586  * @start:  The first block to remove
2587  * @end:   The last block to remove
2588  */
2589 static int
2590 ext4_ext_rm_leaf(handle_t *handle, struct inode *inode,
2591 		 struct ext4_ext_path *path,
2592 		 long long *partial_cluster,
2593 		 ext4_lblk_t start, ext4_lblk_t end)
2594 {
2595 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2596 	int err = 0, correct_index = 0;
2597 	int depth = ext_depth(inode), credits;
2598 	struct ext4_extent_header *eh;
2599 	ext4_lblk_t a, b;
2600 	unsigned num;
2601 	ext4_lblk_t ex_ee_block;
2602 	unsigned short ex_ee_len;
2603 	unsigned unwritten = 0;
2604 	struct ext4_extent *ex;
2605 	ext4_fsblk_t pblk;
2606 
2607 	/* the header must be checked already in ext4_ext_remove_space() */
2608 	ext_debug("truncate since %u in leaf to %u\n", start, end);
2609 	if (!path[depth].p_hdr)
2610 		path[depth].p_hdr = ext_block_hdr(path[depth].p_bh);
2611 	eh = path[depth].p_hdr;
2612 	if (unlikely(path[depth].p_hdr == NULL)) {
2613 		EXT4_ERROR_INODE(inode, "path[%d].p_hdr == NULL", depth);
2614 		return -EIO;
2615 	}
2616 	/* find where to start removing */
2617 	ex = path[depth].p_ext;
2618 	if (!ex)
2619 		ex = EXT_LAST_EXTENT(eh);
2620 
2621 	ex_ee_block = le32_to_cpu(ex->ee_block);
2622 	ex_ee_len = ext4_ext_get_actual_len(ex);
2623 
2624 	/*
2625 	 * If we're starting with an extent other than the last one in the
2626 	 * node, we need to see if it shares a cluster with the extent to
2627 	 * the right (towards the end of the file). If its leftmost cluster
2628 	 * is this extent's rightmost cluster and it is not cluster aligned,
2629 	 * we'll mark it as a partial that is not to be deallocated.
2630 	 */
2631 
2632 	if (ex != EXT_LAST_EXTENT(eh)) {
2633 		ext4_fsblk_t current_pblk, right_pblk;
2634 		long long current_cluster, right_cluster;
2635 
2636 		current_pblk = ext4_ext_pblock(ex) + ex_ee_len - 1;
2637 		current_cluster = (long long)EXT4_B2C(sbi, current_pblk);
2638 		right_pblk = ext4_ext_pblock(ex + 1);
2639 		right_cluster = (long long)EXT4_B2C(sbi, right_pblk);
2640 		if (current_cluster == right_cluster &&
2641 			EXT4_PBLK_COFF(sbi, right_pblk))
2642 			*partial_cluster = -right_cluster;
2643 	}
2644 
2645 	trace_ext4_ext_rm_leaf(inode, start, ex, *partial_cluster);
2646 
2647 	while (ex >= EXT_FIRST_EXTENT(eh) &&
2648 			ex_ee_block + ex_ee_len > start) {
2649 
2650 		if (ext4_ext_is_unwritten(ex))
2651 			unwritten = 1;
2652 		else
2653 			unwritten = 0;
2654 
2655 		ext_debug("remove ext %u:[%d]%d\n", ex_ee_block,
2656 			  unwritten, ex_ee_len);
2657 		path[depth].p_ext = ex;
2658 
2659 		a = ex_ee_block > start ? ex_ee_block : start;
2660 		b = ex_ee_block+ex_ee_len - 1 < end ?
2661 			ex_ee_block+ex_ee_len - 1 : end;
2662 
2663 		ext_debug("  border %u:%u\n", a, b);
2664 
2665 		/* If this extent is beyond the end of the hole, skip it */
2666 		if (end < ex_ee_block) {
2667 			/*
2668 			 * We're going to skip this extent and move to another,
2669 			 * so if this extent is not cluster aligned we have
2670 			 * to mark the current cluster as used to avoid
2671 			 * accidentally freeing it later on
2672 			 */
2673 			pblk = ext4_ext_pblock(ex);
2674 			if (EXT4_PBLK_COFF(sbi, pblk))
2675 				*partial_cluster =
2676 					-((long long)EXT4_B2C(sbi, pblk));
2677 			ex--;
2678 			ex_ee_block = le32_to_cpu(ex->ee_block);
2679 			ex_ee_len = ext4_ext_get_actual_len(ex);
2680 			continue;
2681 		} else if (b != ex_ee_block + ex_ee_len - 1) {
2682 			EXT4_ERROR_INODE(inode,
2683 					 "can not handle truncate %u:%u "
2684 					 "on extent %u:%u",
2685 					 start, end, ex_ee_block,
2686 					 ex_ee_block + ex_ee_len - 1);
2687 			err = -EIO;
2688 			goto out;
2689 		} else if (a != ex_ee_block) {
2690 			/* remove tail of the extent */
2691 			num = a - ex_ee_block;
2692 		} else {
2693 			/* remove whole extent: excellent! */
2694 			num = 0;
2695 		}
2696 		/*
2697 		 * 3 for leaf, sb, and inode plus 2 (bmap and group
2698 		 * descriptor) for each block group; assume two block
2699 		 * groups plus ex_ee_len/blocks_per_block_group for
2700 		 * the worst case
2701 		 */
2702 		credits = 7 + 2*(ex_ee_len/EXT4_BLOCKS_PER_GROUP(inode->i_sb));
2703 		if (ex == EXT_FIRST_EXTENT(eh)) {
2704 			correct_index = 1;
2705 			credits += (ext_depth(inode)) + 1;
2706 		}
2707 		credits += EXT4_MAXQUOTAS_TRANS_BLOCKS(inode->i_sb);
2708 
2709 		err = ext4_ext_truncate_extend_restart(handle, inode, credits);
2710 		if (err)
2711 			goto out;
2712 
2713 		err = ext4_ext_get_access(handle, inode, path + depth);
2714 		if (err)
2715 			goto out;
2716 
2717 		err = ext4_remove_blocks(handle, inode, ex, partial_cluster,
2718 					 a, b);
2719 		if (err)
2720 			goto out;
2721 
2722 		if (num == 0)
2723 			/* this extent is removed; mark slot entirely unused */
2724 			ext4_ext_store_pblock(ex, 0);
2725 
2726 		ex->ee_len = cpu_to_le16(num);
2727 		/*
2728 		 * Do not mark unwritten if all the blocks in the
2729 		 * extent have been removed.
2730 		 */
2731 		if (unwritten && num)
2732 			ext4_ext_mark_unwritten(ex);
2733 		/*
2734 		 * If the extent was completely released,
2735 		 * we need to remove it from the leaf
2736 		 */
2737 		if (num == 0) {
2738 			if (end != EXT_MAX_BLOCKS - 1) {
2739 				/*
2740 				 * For hole punching, we need to scoot all the
2741 				 * extents up when an extent is removed so that
2742 				 * we dont have blank extents in the middle
2743 				 */
2744 				memmove(ex, ex+1, (EXT_LAST_EXTENT(eh) - ex) *
2745 					sizeof(struct ext4_extent));
2746 
2747 				/* Now get rid of the one at the end */
2748 				memset(EXT_LAST_EXTENT(eh), 0,
2749 					sizeof(struct ext4_extent));
2750 			}
2751 			le16_add_cpu(&eh->eh_entries, -1);
2752 		} else if (*partial_cluster > 0)
2753 			*partial_cluster = 0;
2754 
2755 		err = ext4_ext_dirty(handle, inode, path + depth);
2756 		if (err)
2757 			goto out;
2758 
2759 		ext_debug("new extent: %u:%u:%llu\n", ex_ee_block, num,
2760 				ext4_ext_pblock(ex));
2761 		ex--;
2762 		ex_ee_block = le32_to_cpu(ex->ee_block);
2763 		ex_ee_len = ext4_ext_get_actual_len(ex);
2764 	}
2765 
2766 	if (correct_index && eh->eh_entries)
2767 		err = ext4_ext_correct_indexes(handle, inode, path);
2768 
2769 	/*
2770 	 * If there's a partial cluster and at least one extent remains in
2771 	 * the leaf, free the partial cluster if it isn't shared with the
2772 	 * current extent.  If there's a partial cluster and no extents
2773 	 * remain in the leaf, it can't be freed here.  It can only be
2774 	 * freed when it's possible to determine if it's not shared with
2775 	 * any other extent - when the next leaf is processed or when space
2776 	 * removal is complete.
2777 	 */
2778 	if (*partial_cluster > 0 && eh->eh_entries &&
2779 	    (EXT4_B2C(sbi, ext4_ext_pblock(ex) + ex_ee_len - 1) !=
2780 	     *partial_cluster)) {
2781 		int flags = get_default_free_blocks_flags(inode);
2782 
2783 		ext4_free_blocks(handle, inode, NULL,
2784 				 EXT4_C2B(sbi, *partial_cluster),
2785 				 sbi->s_cluster_ratio, flags);
2786 		*partial_cluster = 0;
2787 	}
2788 
2789 	/* if this leaf is free, then we should
2790 	 * remove it from index block above */
2791 	if (err == 0 && eh->eh_entries == 0 && path[depth].p_bh != NULL)
2792 		err = ext4_ext_rm_idx(handle, inode, path, depth);
2793 
2794 out:
2795 	return err;
2796 }
2797 
2798 /*
2799  * ext4_ext_more_to_rm:
2800  * returns 1 if current index has to be freed (even partial)
2801  */
2802 static int
2803 ext4_ext_more_to_rm(struct ext4_ext_path *path)
2804 {
2805 	BUG_ON(path->p_idx == NULL);
2806 
2807 	if (path->p_idx < EXT_FIRST_INDEX(path->p_hdr))
2808 		return 0;
2809 
2810 	/*
2811 	 * if truncate on deeper level happened, it wasn't partial,
2812 	 * so we have to consider current index for truncation
2813 	 */
2814 	if (le16_to_cpu(path->p_hdr->eh_entries) == path->p_block)
2815 		return 0;
2816 	return 1;
2817 }
2818 
2819 int ext4_ext_remove_space(struct inode *inode, ext4_lblk_t start,
2820 			  ext4_lblk_t end)
2821 {
2822 	struct super_block *sb = inode->i_sb;
2823 	int depth = ext_depth(inode);
2824 	struct ext4_ext_path *path = NULL;
2825 	long long partial_cluster = 0;
2826 	handle_t *handle;
2827 	int i = 0, err = 0;
2828 
2829 	ext_debug("truncate since %u to %u\n", start, end);
2830 
2831 	/* probably first extent we're gonna free will be last in block */
2832 	handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, depth + 1);
2833 	if (IS_ERR(handle))
2834 		return PTR_ERR(handle);
2835 
2836 again:
2837 	trace_ext4_ext_remove_space(inode, start, end, depth);
2838 
2839 	/*
2840 	 * Check if we are removing extents inside the extent tree. If that
2841 	 * is the case, we are going to punch a hole inside the extent tree
2842 	 * so we have to check whether we need to split the extent covering
2843 	 * the last block to remove so we can easily remove the part of it
2844 	 * in ext4_ext_rm_leaf().
2845 	 */
2846 	if (end < EXT_MAX_BLOCKS - 1) {
2847 		struct ext4_extent *ex;
2848 		ext4_lblk_t ee_block;
2849 
2850 		/* find extent for this block */
2851 		path = ext4_find_extent(inode, end, NULL, EXT4_EX_NOCACHE);
2852 		if (IS_ERR(path)) {
2853 			ext4_journal_stop(handle);
2854 			return PTR_ERR(path);
2855 		}
2856 		depth = ext_depth(inode);
2857 		/* Leaf not may not exist only if inode has no blocks at all */
2858 		ex = path[depth].p_ext;
2859 		if (!ex) {
2860 			if (depth) {
2861 				EXT4_ERROR_INODE(inode,
2862 						 "path[%d].p_hdr == NULL",
2863 						 depth);
2864 				err = -EIO;
2865 			}
2866 			goto out;
2867 		}
2868 
2869 		ee_block = le32_to_cpu(ex->ee_block);
2870 
2871 		/*
2872 		 * See if the last block is inside the extent, if so split
2873 		 * the extent at 'end' block so we can easily remove the
2874 		 * tail of the first part of the split extent in
2875 		 * ext4_ext_rm_leaf().
2876 		 */
2877 		if (end >= ee_block &&
2878 		    end < ee_block + ext4_ext_get_actual_len(ex) - 1) {
2879 			/*
2880 			 * Split the extent in two so that 'end' is the last
2881 			 * block in the first new extent. Also we should not
2882 			 * fail removing space due to ENOSPC so try to use
2883 			 * reserved block if that happens.
2884 			 */
2885 			err = ext4_force_split_extent_at(handle, inode, &path,
2886 							 end + 1, 1);
2887 			if (err < 0)
2888 				goto out;
2889 		}
2890 	}
2891 	/*
2892 	 * We start scanning from right side, freeing all the blocks
2893 	 * after i_size and walking into the tree depth-wise.
2894 	 */
2895 	depth = ext_depth(inode);
2896 	if (path) {
2897 		int k = i = depth;
2898 		while (--k > 0)
2899 			path[k].p_block =
2900 				le16_to_cpu(path[k].p_hdr->eh_entries)+1;
2901 	} else {
2902 		path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
2903 			       GFP_NOFS);
2904 		if (path == NULL) {
2905 			ext4_journal_stop(handle);
2906 			return -ENOMEM;
2907 		}
2908 		path[0].p_maxdepth = path[0].p_depth = depth;
2909 		path[0].p_hdr = ext_inode_hdr(inode);
2910 		i = 0;
2911 
2912 		if (ext4_ext_check(inode, path[0].p_hdr, depth, 0)) {
2913 			err = -EIO;
2914 			goto out;
2915 		}
2916 	}
2917 	err = 0;
2918 
2919 	while (i >= 0 && err == 0) {
2920 		if (i == depth) {
2921 			/* this is leaf block */
2922 			err = ext4_ext_rm_leaf(handle, inode, path,
2923 					       &partial_cluster, start,
2924 					       end);
2925 			/* root level has p_bh == NULL, brelse() eats this */
2926 			brelse(path[i].p_bh);
2927 			path[i].p_bh = NULL;
2928 			i--;
2929 			continue;
2930 		}
2931 
2932 		/* this is index block */
2933 		if (!path[i].p_hdr) {
2934 			ext_debug("initialize header\n");
2935 			path[i].p_hdr = ext_block_hdr(path[i].p_bh);
2936 		}
2937 
2938 		if (!path[i].p_idx) {
2939 			/* this level hasn't been touched yet */
2940 			path[i].p_idx = EXT_LAST_INDEX(path[i].p_hdr);
2941 			path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries)+1;
2942 			ext_debug("init index ptr: hdr 0x%p, num %d\n",
2943 				  path[i].p_hdr,
2944 				  le16_to_cpu(path[i].p_hdr->eh_entries));
2945 		} else {
2946 			/* we were already here, see at next index */
2947 			path[i].p_idx--;
2948 		}
2949 
2950 		ext_debug("level %d - index, first 0x%p, cur 0x%p\n",
2951 				i, EXT_FIRST_INDEX(path[i].p_hdr),
2952 				path[i].p_idx);
2953 		if (ext4_ext_more_to_rm(path + i)) {
2954 			struct buffer_head *bh;
2955 			/* go to the next level */
2956 			ext_debug("move to level %d (block %llu)\n",
2957 				  i + 1, ext4_idx_pblock(path[i].p_idx));
2958 			memset(path + i + 1, 0, sizeof(*path));
2959 			bh = read_extent_tree_block(inode,
2960 				ext4_idx_pblock(path[i].p_idx), depth - i - 1,
2961 				EXT4_EX_NOCACHE);
2962 			if (IS_ERR(bh)) {
2963 				/* should we reset i_size? */
2964 				err = PTR_ERR(bh);
2965 				break;
2966 			}
2967 			/* Yield here to deal with large extent trees.
2968 			 * Should be a no-op if we did IO above. */
2969 			cond_resched();
2970 			if (WARN_ON(i + 1 > depth)) {
2971 				err = -EIO;
2972 				break;
2973 			}
2974 			path[i + 1].p_bh = bh;
2975 
2976 			/* save actual number of indexes since this
2977 			 * number is changed at the next iteration */
2978 			path[i].p_block = le16_to_cpu(path[i].p_hdr->eh_entries);
2979 			i++;
2980 		} else {
2981 			/* we finished processing this index, go up */
2982 			if (path[i].p_hdr->eh_entries == 0 && i > 0) {
2983 				/* index is empty, remove it;
2984 				 * handle must be already prepared by the
2985 				 * truncatei_leaf() */
2986 				err = ext4_ext_rm_idx(handle, inode, path, i);
2987 			}
2988 			/* root level has p_bh == NULL, brelse() eats this */
2989 			brelse(path[i].p_bh);
2990 			path[i].p_bh = NULL;
2991 			i--;
2992 			ext_debug("return to level %d\n", i);
2993 		}
2994 	}
2995 
2996 	trace_ext4_ext_remove_space_done(inode, start, end, depth,
2997 			partial_cluster, path->p_hdr->eh_entries);
2998 
2999 	/* If we still have something in the partial cluster and we have removed
3000 	 * even the first extent, then we should free the blocks in the partial
3001 	 * cluster as well. */
3002 	if (partial_cluster > 0 && path->p_hdr->eh_entries == 0) {
3003 		int flags = get_default_free_blocks_flags(inode);
3004 
3005 		ext4_free_blocks(handle, inode, NULL,
3006 				 EXT4_C2B(EXT4_SB(sb), partial_cluster),
3007 				 EXT4_SB(sb)->s_cluster_ratio, flags);
3008 		partial_cluster = 0;
3009 	}
3010 
3011 	/* TODO: flexible tree reduction should be here */
3012 	if (path->p_hdr->eh_entries == 0) {
3013 		/*
3014 		 * truncate to zero freed all the tree,
3015 		 * so we need to correct eh_depth
3016 		 */
3017 		err = ext4_ext_get_access(handle, inode, path);
3018 		if (err == 0) {
3019 			ext_inode_hdr(inode)->eh_depth = 0;
3020 			ext_inode_hdr(inode)->eh_max =
3021 				cpu_to_le16(ext4_ext_space_root(inode, 0));
3022 			err = ext4_ext_dirty(handle, inode, path);
3023 		}
3024 	}
3025 out:
3026 	ext4_ext_drop_refs(path);
3027 	kfree(path);
3028 	path = NULL;
3029 	if (err == -EAGAIN)
3030 		goto again;
3031 	ext4_journal_stop(handle);
3032 
3033 	return err;
3034 }
3035 
3036 /*
3037  * called at mount time
3038  */
3039 void ext4_ext_init(struct super_block *sb)
3040 {
3041 	/*
3042 	 * possible initialization would be here
3043 	 */
3044 
3045 	if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS)) {
3046 #if defined(AGGRESSIVE_TEST) || defined(CHECK_BINSEARCH) || defined(EXTENTS_STATS)
3047 		printk(KERN_INFO "EXT4-fs: file extents enabled"
3048 #ifdef AGGRESSIVE_TEST
3049 		       ", aggressive tests"
3050 #endif
3051 #ifdef CHECK_BINSEARCH
3052 		       ", check binsearch"
3053 #endif
3054 #ifdef EXTENTS_STATS
3055 		       ", stats"
3056 #endif
3057 		       "\n");
3058 #endif
3059 #ifdef EXTENTS_STATS
3060 		spin_lock_init(&EXT4_SB(sb)->s_ext_stats_lock);
3061 		EXT4_SB(sb)->s_ext_min = 1 << 30;
3062 		EXT4_SB(sb)->s_ext_max = 0;
3063 #endif
3064 	}
3065 }
3066 
3067 /*
3068  * called at umount time
3069  */
3070 void ext4_ext_release(struct super_block *sb)
3071 {
3072 	if (!EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EXTENTS))
3073 		return;
3074 
3075 #ifdef EXTENTS_STATS
3076 	if (EXT4_SB(sb)->s_ext_blocks && EXT4_SB(sb)->s_ext_extents) {
3077 		struct ext4_sb_info *sbi = EXT4_SB(sb);
3078 		printk(KERN_ERR "EXT4-fs: %lu blocks in %lu extents (%lu ave)\n",
3079 			sbi->s_ext_blocks, sbi->s_ext_extents,
3080 			sbi->s_ext_blocks / sbi->s_ext_extents);
3081 		printk(KERN_ERR "EXT4-fs: extents: %lu min, %lu max, max depth %lu\n",
3082 			sbi->s_ext_min, sbi->s_ext_max, sbi->s_depth_max);
3083 	}
3084 #endif
3085 }
3086 
3087 static int ext4_zeroout_es(struct inode *inode, struct ext4_extent *ex)
3088 {
3089 	ext4_lblk_t  ee_block;
3090 	ext4_fsblk_t ee_pblock;
3091 	unsigned int ee_len;
3092 
3093 	ee_block  = le32_to_cpu(ex->ee_block);
3094 	ee_len    = ext4_ext_get_actual_len(ex);
3095 	ee_pblock = ext4_ext_pblock(ex);
3096 
3097 	if (ee_len == 0)
3098 		return 0;
3099 
3100 	return ext4_es_insert_extent(inode, ee_block, ee_len, ee_pblock,
3101 				     EXTENT_STATUS_WRITTEN);
3102 }
3103 
3104 /* FIXME!! we need to try to merge to left or right after zero-out  */
3105 static int ext4_ext_zeroout(struct inode *inode, struct ext4_extent *ex)
3106 {
3107 	ext4_fsblk_t ee_pblock;
3108 	unsigned int ee_len;
3109 	int ret;
3110 
3111 	ee_len    = ext4_ext_get_actual_len(ex);
3112 	ee_pblock = ext4_ext_pblock(ex);
3113 
3114 	ret = sb_issue_zeroout(inode->i_sb, ee_pblock, ee_len, GFP_NOFS);
3115 	if (ret > 0)
3116 		ret = 0;
3117 
3118 	return ret;
3119 }
3120 
3121 /*
3122  * ext4_split_extent_at() splits an extent at given block.
3123  *
3124  * @handle: the journal handle
3125  * @inode: the file inode
3126  * @path: the path to the extent
3127  * @split: the logical block where the extent is splitted.
3128  * @split_flags: indicates if the extent could be zeroout if split fails, and
3129  *		 the states(init or unwritten) of new extents.
3130  * @flags: flags used to insert new extent to extent tree.
3131  *
3132  *
3133  * Splits extent [a, b] into two extents [a, @split) and [@split, b], states
3134  * of which are deterimined by split_flag.
3135  *
3136  * There are two cases:
3137  *  a> the extent are splitted into two extent.
3138  *  b> split is not needed, and just mark the extent.
3139  *
3140  * return 0 on success.
3141  */
3142 static int ext4_split_extent_at(handle_t *handle,
3143 			     struct inode *inode,
3144 			     struct ext4_ext_path **ppath,
3145 			     ext4_lblk_t split,
3146 			     int split_flag,
3147 			     int flags)
3148 {
3149 	struct ext4_ext_path *path = *ppath;
3150 	ext4_fsblk_t newblock;
3151 	ext4_lblk_t ee_block;
3152 	struct ext4_extent *ex, newex, orig_ex, zero_ex;
3153 	struct ext4_extent *ex2 = NULL;
3154 	unsigned int ee_len, depth;
3155 	int err = 0;
3156 
3157 	BUG_ON((split_flag & (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2)) ==
3158 	       (EXT4_EXT_DATA_VALID1 | EXT4_EXT_DATA_VALID2));
3159 
3160 	ext_debug("ext4_split_extents_at: inode %lu, logical"
3161 		"block %llu\n", inode->i_ino, (unsigned long long)split);
3162 
3163 	ext4_ext_show_leaf(inode, path);
3164 
3165 	depth = ext_depth(inode);
3166 	ex = path[depth].p_ext;
3167 	ee_block = le32_to_cpu(ex->ee_block);
3168 	ee_len = ext4_ext_get_actual_len(ex);
3169 	newblock = split - ee_block + ext4_ext_pblock(ex);
3170 
3171 	BUG_ON(split < ee_block || split >= (ee_block + ee_len));
3172 	BUG_ON(!ext4_ext_is_unwritten(ex) &&
3173 	       split_flag & (EXT4_EXT_MAY_ZEROOUT |
3174 			     EXT4_EXT_MARK_UNWRIT1 |
3175 			     EXT4_EXT_MARK_UNWRIT2));
3176 
3177 	err = ext4_ext_get_access(handle, inode, path + depth);
3178 	if (err)
3179 		goto out;
3180 
3181 	if (split == ee_block) {
3182 		/*
3183 		 * case b: block @split is the block that the extent begins with
3184 		 * then we just change the state of the extent, and splitting
3185 		 * is not needed.
3186 		 */
3187 		if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3188 			ext4_ext_mark_unwritten(ex);
3189 		else
3190 			ext4_ext_mark_initialized(ex);
3191 
3192 		if (!(flags & EXT4_GET_BLOCKS_PRE_IO))
3193 			ext4_ext_try_to_merge(handle, inode, path, ex);
3194 
3195 		err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3196 		goto out;
3197 	}
3198 
3199 	/* case a */
3200 	memcpy(&orig_ex, ex, sizeof(orig_ex));
3201 	ex->ee_len = cpu_to_le16(split - ee_block);
3202 	if (split_flag & EXT4_EXT_MARK_UNWRIT1)
3203 		ext4_ext_mark_unwritten(ex);
3204 
3205 	/*
3206 	 * path may lead to new leaf, not to original leaf any more
3207 	 * after ext4_ext_insert_extent() returns,
3208 	 */
3209 	err = ext4_ext_dirty(handle, inode, path + depth);
3210 	if (err)
3211 		goto fix_extent_len;
3212 
3213 	ex2 = &newex;
3214 	ex2->ee_block = cpu_to_le32(split);
3215 	ex2->ee_len   = cpu_to_le16(ee_len - (split - ee_block));
3216 	ext4_ext_store_pblock(ex2, newblock);
3217 	if (split_flag & EXT4_EXT_MARK_UNWRIT2)
3218 		ext4_ext_mark_unwritten(ex2);
3219 
3220 	err = ext4_ext_insert_extent(handle, inode, ppath, &newex, flags);
3221 	if (err == -ENOSPC && (EXT4_EXT_MAY_ZEROOUT & split_flag)) {
3222 		if (split_flag & (EXT4_EXT_DATA_VALID1|EXT4_EXT_DATA_VALID2)) {
3223 			if (split_flag & EXT4_EXT_DATA_VALID1) {
3224 				err = ext4_ext_zeroout(inode, ex2);
3225 				zero_ex.ee_block = ex2->ee_block;
3226 				zero_ex.ee_len = cpu_to_le16(
3227 						ext4_ext_get_actual_len(ex2));
3228 				ext4_ext_store_pblock(&zero_ex,
3229 						      ext4_ext_pblock(ex2));
3230 			} else {
3231 				err = ext4_ext_zeroout(inode, ex);
3232 				zero_ex.ee_block = ex->ee_block;
3233 				zero_ex.ee_len = cpu_to_le16(
3234 						ext4_ext_get_actual_len(ex));
3235 				ext4_ext_store_pblock(&zero_ex,
3236 						      ext4_ext_pblock(ex));
3237 			}
3238 		} else {
3239 			err = ext4_ext_zeroout(inode, &orig_ex);
3240 			zero_ex.ee_block = orig_ex.ee_block;
3241 			zero_ex.ee_len = cpu_to_le16(
3242 						ext4_ext_get_actual_len(&orig_ex));
3243 			ext4_ext_store_pblock(&zero_ex,
3244 					      ext4_ext_pblock(&orig_ex));
3245 		}
3246 
3247 		if (err)
3248 			goto fix_extent_len;
3249 		/* update the extent length and mark as initialized */
3250 		ex->ee_len = cpu_to_le16(ee_len);
3251 		ext4_ext_try_to_merge(handle, inode, path, ex);
3252 		err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3253 		if (err)
3254 			goto fix_extent_len;
3255 
3256 		/* update extent status tree */
3257 		err = ext4_zeroout_es(inode, &zero_ex);
3258 
3259 		goto out;
3260 	} else if (err)
3261 		goto fix_extent_len;
3262 
3263 out:
3264 	ext4_ext_show_leaf(inode, path);
3265 	return err;
3266 
3267 fix_extent_len:
3268 	ex->ee_len = orig_ex.ee_len;
3269 	ext4_ext_dirty(handle, inode, path + path->p_depth);
3270 	return err;
3271 }
3272 
3273 /*
3274  * ext4_split_extents() splits an extent and mark extent which is covered
3275  * by @map as split_flags indicates
3276  *
3277  * It may result in splitting the extent into multiple extents (up to three)
3278  * There are three possibilities:
3279  *   a> There is no split required
3280  *   b> Splits in two extents: Split is happening at either end of the extent
3281  *   c> Splits in three extents: Somone is splitting in middle of the extent
3282  *
3283  */
3284 static int ext4_split_extent(handle_t *handle,
3285 			      struct inode *inode,
3286 			      struct ext4_ext_path **ppath,
3287 			      struct ext4_map_blocks *map,
3288 			      int split_flag,
3289 			      int flags)
3290 {
3291 	struct ext4_ext_path *path = *ppath;
3292 	ext4_lblk_t ee_block;
3293 	struct ext4_extent *ex;
3294 	unsigned int ee_len, depth;
3295 	int err = 0;
3296 	int unwritten;
3297 	int split_flag1, flags1;
3298 	int allocated = map->m_len;
3299 
3300 	depth = ext_depth(inode);
3301 	ex = path[depth].p_ext;
3302 	ee_block = le32_to_cpu(ex->ee_block);
3303 	ee_len = ext4_ext_get_actual_len(ex);
3304 	unwritten = ext4_ext_is_unwritten(ex);
3305 
3306 	if (map->m_lblk + map->m_len < ee_block + ee_len) {
3307 		split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
3308 		flags1 = flags | EXT4_GET_BLOCKS_PRE_IO;
3309 		if (unwritten)
3310 			split_flag1 |= EXT4_EXT_MARK_UNWRIT1 |
3311 				       EXT4_EXT_MARK_UNWRIT2;
3312 		if (split_flag & EXT4_EXT_DATA_VALID2)
3313 			split_flag1 |= EXT4_EXT_DATA_VALID1;
3314 		err = ext4_split_extent_at(handle, inode, ppath,
3315 				map->m_lblk + map->m_len, split_flag1, flags1);
3316 		if (err)
3317 			goto out;
3318 	} else {
3319 		allocated = ee_len - (map->m_lblk - ee_block);
3320 	}
3321 	/*
3322 	 * Update path is required because previous ext4_split_extent_at() may
3323 	 * result in split of original leaf or extent zeroout.
3324 	 */
3325 	path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3326 	if (IS_ERR(path))
3327 		return PTR_ERR(path);
3328 	depth = ext_depth(inode);
3329 	ex = path[depth].p_ext;
3330 	if (!ex) {
3331 		EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3332 				 (unsigned long) map->m_lblk);
3333 		return -EIO;
3334 	}
3335 	unwritten = ext4_ext_is_unwritten(ex);
3336 	split_flag1 = 0;
3337 
3338 	if (map->m_lblk >= ee_block) {
3339 		split_flag1 = split_flag & EXT4_EXT_DATA_VALID2;
3340 		if (unwritten) {
3341 			split_flag1 |= EXT4_EXT_MARK_UNWRIT1;
3342 			split_flag1 |= split_flag & (EXT4_EXT_MAY_ZEROOUT |
3343 						     EXT4_EXT_MARK_UNWRIT2);
3344 		}
3345 		err = ext4_split_extent_at(handle, inode, ppath,
3346 				map->m_lblk, split_flag1, flags);
3347 		if (err)
3348 			goto out;
3349 	}
3350 
3351 	ext4_ext_show_leaf(inode, path);
3352 out:
3353 	return err ? err : allocated;
3354 }
3355 
3356 /*
3357  * This function is called by ext4_ext_map_blocks() if someone tries to write
3358  * to an unwritten extent. It may result in splitting the unwritten
3359  * extent into multiple extents (up to three - one initialized and two
3360  * unwritten).
3361  * There are three possibilities:
3362  *   a> There is no split required: Entire extent should be initialized
3363  *   b> Splits in two extents: Write is happening at either end of the extent
3364  *   c> Splits in three extents: Somone is writing in middle of the extent
3365  *
3366  * Pre-conditions:
3367  *  - The extent pointed to by 'path' is unwritten.
3368  *  - The extent pointed to by 'path' contains a superset
3369  *    of the logical span [map->m_lblk, map->m_lblk + map->m_len).
3370  *
3371  * Post-conditions on success:
3372  *  - the returned value is the number of blocks beyond map->l_lblk
3373  *    that are allocated and initialized.
3374  *    It is guaranteed to be >= map->m_len.
3375  */
3376 static int ext4_ext_convert_to_initialized(handle_t *handle,
3377 					   struct inode *inode,
3378 					   struct ext4_map_blocks *map,
3379 					   struct ext4_ext_path **ppath,
3380 					   int flags)
3381 {
3382 	struct ext4_ext_path *path = *ppath;
3383 	struct ext4_sb_info *sbi;
3384 	struct ext4_extent_header *eh;
3385 	struct ext4_map_blocks split_map;
3386 	struct ext4_extent zero_ex;
3387 	struct ext4_extent *ex, *abut_ex;
3388 	ext4_lblk_t ee_block, eof_block;
3389 	unsigned int ee_len, depth, map_len = map->m_len;
3390 	int allocated = 0, max_zeroout = 0;
3391 	int err = 0;
3392 	int split_flag = 0;
3393 
3394 	ext_debug("ext4_ext_convert_to_initialized: inode %lu, logical"
3395 		"block %llu, max_blocks %u\n", inode->i_ino,
3396 		(unsigned long long)map->m_lblk, map_len);
3397 
3398 	sbi = EXT4_SB(inode->i_sb);
3399 	eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3400 		inode->i_sb->s_blocksize_bits;
3401 	if (eof_block < map->m_lblk + map_len)
3402 		eof_block = map->m_lblk + map_len;
3403 
3404 	depth = ext_depth(inode);
3405 	eh = path[depth].p_hdr;
3406 	ex = path[depth].p_ext;
3407 	ee_block = le32_to_cpu(ex->ee_block);
3408 	ee_len = ext4_ext_get_actual_len(ex);
3409 	zero_ex.ee_len = 0;
3410 
3411 	trace_ext4_ext_convert_to_initialized_enter(inode, map, ex);
3412 
3413 	/* Pre-conditions */
3414 	BUG_ON(!ext4_ext_is_unwritten(ex));
3415 	BUG_ON(!in_range(map->m_lblk, ee_block, ee_len));
3416 
3417 	/*
3418 	 * Attempt to transfer newly initialized blocks from the currently
3419 	 * unwritten extent to its neighbor. This is much cheaper
3420 	 * than an insertion followed by a merge as those involve costly
3421 	 * memmove() calls. Transferring to the left is the common case in
3422 	 * steady state for workloads doing fallocate(FALLOC_FL_KEEP_SIZE)
3423 	 * followed by append writes.
3424 	 *
3425 	 * Limitations of the current logic:
3426 	 *  - L1: we do not deal with writes covering the whole extent.
3427 	 *    This would require removing the extent if the transfer
3428 	 *    is possible.
3429 	 *  - L2: we only attempt to merge with an extent stored in the
3430 	 *    same extent tree node.
3431 	 */
3432 	if ((map->m_lblk == ee_block) &&
3433 		/* See if we can merge left */
3434 		(map_len < ee_len) &&		/*L1*/
3435 		(ex > EXT_FIRST_EXTENT(eh))) {	/*L2*/
3436 		ext4_lblk_t prev_lblk;
3437 		ext4_fsblk_t prev_pblk, ee_pblk;
3438 		unsigned int prev_len;
3439 
3440 		abut_ex = ex - 1;
3441 		prev_lblk = le32_to_cpu(abut_ex->ee_block);
3442 		prev_len = ext4_ext_get_actual_len(abut_ex);
3443 		prev_pblk = ext4_ext_pblock(abut_ex);
3444 		ee_pblk = ext4_ext_pblock(ex);
3445 
3446 		/*
3447 		 * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3448 		 * upon those conditions:
3449 		 * - C1: abut_ex is initialized,
3450 		 * - C2: abut_ex is logically abutting ex,
3451 		 * - C3: abut_ex is physically abutting ex,
3452 		 * - C4: abut_ex can receive the additional blocks without
3453 		 *   overflowing the (initialized) length limit.
3454 		 */
3455 		if ((!ext4_ext_is_unwritten(abut_ex)) &&		/*C1*/
3456 			((prev_lblk + prev_len) == ee_block) &&		/*C2*/
3457 			((prev_pblk + prev_len) == ee_pblk) &&		/*C3*/
3458 			(prev_len < (EXT_INIT_MAX_LEN - map_len))) {	/*C4*/
3459 			err = ext4_ext_get_access(handle, inode, path + depth);
3460 			if (err)
3461 				goto out;
3462 
3463 			trace_ext4_ext_convert_to_initialized_fastpath(inode,
3464 				map, ex, abut_ex);
3465 
3466 			/* Shift the start of ex by 'map_len' blocks */
3467 			ex->ee_block = cpu_to_le32(ee_block + map_len);
3468 			ext4_ext_store_pblock(ex, ee_pblk + map_len);
3469 			ex->ee_len = cpu_to_le16(ee_len - map_len);
3470 			ext4_ext_mark_unwritten(ex); /* Restore the flag */
3471 
3472 			/* Extend abut_ex by 'map_len' blocks */
3473 			abut_ex->ee_len = cpu_to_le16(prev_len + map_len);
3474 
3475 			/* Result: number of initialized blocks past m_lblk */
3476 			allocated = map_len;
3477 		}
3478 	} else if (((map->m_lblk + map_len) == (ee_block + ee_len)) &&
3479 		   (map_len < ee_len) &&	/*L1*/
3480 		   ex < EXT_LAST_EXTENT(eh)) {	/*L2*/
3481 		/* See if we can merge right */
3482 		ext4_lblk_t next_lblk;
3483 		ext4_fsblk_t next_pblk, ee_pblk;
3484 		unsigned int next_len;
3485 
3486 		abut_ex = ex + 1;
3487 		next_lblk = le32_to_cpu(abut_ex->ee_block);
3488 		next_len = ext4_ext_get_actual_len(abut_ex);
3489 		next_pblk = ext4_ext_pblock(abut_ex);
3490 		ee_pblk = ext4_ext_pblock(ex);
3491 
3492 		/*
3493 		 * A transfer of blocks from 'ex' to 'abut_ex' is allowed
3494 		 * upon those conditions:
3495 		 * - C1: abut_ex is initialized,
3496 		 * - C2: abut_ex is logically abutting ex,
3497 		 * - C3: abut_ex is physically abutting ex,
3498 		 * - C4: abut_ex can receive the additional blocks without
3499 		 *   overflowing the (initialized) length limit.
3500 		 */
3501 		if ((!ext4_ext_is_unwritten(abut_ex)) &&		/*C1*/
3502 		    ((map->m_lblk + map_len) == next_lblk) &&		/*C2*/
3503 		    ((ee_pblk + ee_len) == next_pblk) &&		/*C3*/
3504 		    (next_len < (EXT_INIT_MAX_LEN - map_len))) {	/*C4*/
3505 			err = ext4_ext_get_access(handle, inode, path + depth);
3506 			if (err)
3507 				goto out;
3508 
3509 			trace_ext4_ext_convert_to_initialized_fastpath(inode,
3510 				map, ex, abut_ex);
3511 
3512 			/* Shift the start of abut_ex by 'map_len' blocks */
3513 			abut_ex->ee_block = cpu_to_le32(next_lblk - map_len);
3514 			ext4_ext_store_pblock(abut_ex, next_pblk - map_len);
3515 			ex->ee_len = cpu_to_le16(ee_len - map_len);
3516 			ext4_ext_mark_unwritten(ex); /* Restore the flag */
3517 
3518 			/* Extend abut_ex by 'map_len' blocks */
3519 			abut_ex->ee_len = cpu_to_le16(next_len + map_len);
3520 
3521 			/* Result: number of initialized blocks past m_lblk */
3522 			allocated = map_len;
3523 		}
3524 	}
3525 	if (allocated) {
3526 		/* Mark the block containing both extents as dirty */
3527 		ext4_ext_dirty(handle, inode, path + depth);
3528 
3529 		/* Update path to point to the right extent */
3530 		path[depth].p_ext = abut_ex;
3531 		goto out;
3532 	} else
3533 		allocated = ee_len - (map->m_lblk - ee_block);
3534 
3535 	WARN_ON(map->m_lblk < ee_block);
3536 	/*
3537 	 * It is safe to convert extent to initialized via explicit
3538 	 * zeroout only if extent is fully inside i_size or new_size.
3539 	 */
3540 	split_flag |= ee_block + ee_len <= eof_block ? EXT4_EXT_MAY_ZEROOUT : 0;
3541 
3542 	if (EXT4_EXT_MAY_ZEROOUT & split_flag)
3543 		max_zeroout = sbi->s_extent_max_zeroout_kb >>
3544 			(inode->i_sb->s_blocksize_bits - 10);
3545 
3546 	/* If extent is less than s_max_zeroout_kb, zeroout directly */
3547 	if (max_zeroout && (ee_len <= max_zeroout)) {
3548 		err = ext4_ext_zeroout(inode, ex);
3549 		if (err)
3550 			goto out;
3551 		zero_ex.ee_block = ex->ee_block;
3552 		zero_ex.ee_len = cpu_to_le16(ext4_ext_get_actual_len(ex));
3553 		ext4_ext_store_pblock(&zero_ex, ext4_ext_pblock(ex));
3554 
3555 		err = ext4_ext_get_access(handle, inode, path + depth);
3556 		if (err)
3557 			goto out;
3558 		ext4_ext_mark_initialized(ex);
3559 		ext4_ext_try_to_merge(handle, inode, path, ex);
3560 		err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3561 		goto out;
3562 	}
3563 
3564 	/*
3565 	 * four cases:
3566 	 * 1. split the extent into three extents.
3567 	 * 2. split the extent into two extents, zeroout the first half.
3568 	 * 3. split the extent into two extents, zeroout the second half.
3569 	 * 4. split the extent into two extents with out zeroout.
3570 	 */
3571 	split_map.m_lblk = map->m_lblk;
3572 	split_map.m_len = map->m_len;
3573 
3574 	if (max_zeroout && (allocated > map->m_len)) {
3575 		if (allocated <= max_zeroout) {
3576 			/* case 3 */
3577 			zero_ex.ee_block =
3578 					 cpu_to_le32(map->m_lblk);
3579 			zero_ex.ee_len = cpu_to_le16(allocated);
3580 			ext4_ext_store_pblock(&zero_ex,
3581 				ext4_ext_pblock(ex) + map->m_lblk - ee_block);
3582 			err = ext4_ext_zeroout(inode, &zero_ex);
3583 			if (err)
3584 				goto out;
3585 			split_map.m_lblk = map->m_lblk;
3586 			split_map.m_len = allocated;
3587 		} else if (map->m_lblk - ee_block + map->m_len < max_zeroout) {
3588 			/* case 2 */
3589 			if (map->m_lblk != ee_block) {
3590 				zero_ex.ee_block = ex->ee_block;
3591 				zero_ex.ee_len = cpu_to_le16(map->m_lblk -
3592 							ee_block);
3593 				ext4_ext_store_pblock(&zero_ex,
3594 						      ext4_ext_pblock(ex));
3595 				err = ext4_ext_zeroout(inode, &zero_ex);
3596 				if (err)
3597 					goto out;
3598 			}
3599 
3600 			split_map.m_lblk = ee_block;
3601 			split_map.m_len = map->m_lblk - ee_block + map->m_len;
3602 			allocated = map->m_len;
3603 		}
3604 	}
3605 
3606 	err = ext4_split_extent(handle, inode, ppath, &split_map, split_flag,
3607 				flags);
3608 	if (err > 0)
3609 		err = 0;
3610 out:
3611 	/* If we have gotten a failure, don't zero out status tree */
3612 	if (!err)
3613 		err = ext4_zeroout_es(inode, &zero_ex);
3614 	return err ? err : allocated;
3615 }
3616 
3617 /*
3618  * This function is called by ext4_ext_map_blocks() from
3619  * ext4_get_blocks_dio_write() when DIO to write
3620  * to an unwritten extent.
3621  *
3622  * Writing to an unwritten extent may result in splitting the unwritten
3623  * extent into multiple initialized/unwritten extents (up to three)
3624  * There are three possibilities:
3625  *   a> There is no split required: Entire extent should be unwritten
3626  *   b> Splits in two extents: Write is happening at either end of the extent
3627  *   c> Splits in three extents: Somone is writing in middle of the extent
3628  *
3629  * This works the same way in the case of initialized -> unwritten conversion.
3630  *
3631  * One of more index blocks maybe needed if the extent tree grow after
3632  * the unwritten extent split. To prevent ENOSPC occur at the IO
3633  * complete, we need to split the unwritten extent before DIO submit
3634  * the IO. The unwritten extent called at this time will be split
3635  * into three unwritten extent(at most). After IO complete, the part
3636  * being filled will be convert to initialized by the end_io callback function
3637  * via ext4_convert_unwritten_extents().
3638  *
3639  * Returns the size of unwritten extent to be written on success.
3640  */
3641 static int ext4_split_convert_extents(handle_t *handle,
3642 					struct inode *inode,
3643 					struct ext4_map_blocks *map,
3644 					struct ext4_ext_path **ppath,
3645 					int flags)
3646 {
3647 	struct ext4_ext_path *path = *ppath;
3648 	ext4_lblk_t eof_block;
3649 	ext4_lblk_t ee_block;
3650 	struct ext4_extent *ex;
3651 	unsigned int ee_len;
3652 	int split_flag = 0, depth;
3653 
3654 	ext_debug("%s: inode %lu, logical block %llu, max_blocks %u\n",
3655 		  __func__, inode->i_ino,
3656 		  (unsigned long long)map->m_lblk, map->m_len);
3657 
3658 	eof_block = (inode->i_size + inode->i_sb->s_blocksize - 1) >>
3659 		inode->i_sb->s_blocksize_bits;
3660 	if (eof_block < map->m_lblk + map->m_len)
3661 		eof_block = map->m_lblk + map->m_len;
3662 	/*
3663 	 * It is safe to convert extent to initialized via explicit
3664 	 * zeroout only if extent is fully insde i_size or new_size.
3665 	 */
3666 	depth = ext_depth(inode);
3667 	ex = path[depth].p_ext;
3668 	ee_block = le32_to_cpu(ex->ee_block);
3669 	ee_len = ext4_ext_get_actual_len(ex);
3670 
3671 	/* Convert to unwritten */
3672 	if (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN) {
3673 		split_flag |= EXT4_EXT_DATA_VALID1;
3674 	/* Convert to initialized */
3675 	} else if (flags & EXT4_GET_BLOCKS_CONVERT) {
3676 		split_flag |= ee_block + ee_len <= eof_block ?
3677 			      EXT4_EXT_MAY_ZEROOUT : 0;
3678 		split_flag |= (EXT4_EXT_MARK_UNWRIT2 | EXT4_EXT_DATA_VALID2);
3679 	}
3680 	flags |= EXT4_GET_BLOCKS_PRE_IO;
3681 	return ext4_split_extent(handle, inode, ppath, map, split_flag, flags);
3682 }
3683 
3684 static int ext4_convert_unwritten_extents_endio(handle_t *handle,
3685 						struct inode *inode,
3686 						struct ext4_map_blocks *map,
3687 						struct ext4_ext_path **ppath)
3688 {
3689 	struct ext4_ext_path *path = *ppath;
3690 	struct ext4_extent *ex;
3691 	ext4_lblk_t ee_block;
3692 	unsigned int ee_len;
3693 	int depth;
3694 	int err = 0;
3695 
3696 	depth = ext_depth(inode);
3697 	ex = path[depth].p_ext;
3698 	ee_block = le32_to_cpu(ex->ee_block);
3699 	ee_len = ext4_ext_get_actual_len(ex);
3700 
3701 	ext_debug("ext4_convert_unwritten_extents_endio: inode %lu, logical"
3702 		"block %llu, max_blocks %u\n", inode->i_ino,
3703 		  (unsigned long long)ee_block, ee_len);
3704 
3705 	/* If extent is larger than requested it is a clear sign that we still
3706 	 * have some extent state machine issues left. So extent_split is still
3707 	 * required.
3708 	 * TODO: Once all related issues will be fixed this situation should be
3709 	 * illegal.
3710 	 */
3711 	if (ee_block != map->m_lblk || ee_len > map->m_len) {
3712 #ifdef EXT4_DEBUG
3713 		ext4_warning("Inode (%ld) finished: extent logical block %llu,"
3714 			     " len %u; IO logical block %llu, len %u\n",
3715 			     inode->i_ino, (unsigned long long)ee_block, ee_len,
3716 			     (unsigned long long)map->m_lblk, map->m_len);
3717 #endif
3718 		err = ext4_split_convert_extents(handle, inode, map, ppath,
3719 						 EXT4_GET_BLOCKS_CONVERT);
3720 		if (err < 0)
3721 			return err;
3722 		path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3723 		if (IS_ERR(path))
3724 			return PTR_ERR(path);
3725 		depth = ext_depth(inode);
3726 		ex = path[depth].p_ext;
3727 	}
3728 
3729 	err = ext4_ext_get_access(handle, inode, path + depth);
3730 	if (err)
3731 		goto out;
3732 	/* first mark the extent as initialized */
3733 	ext4_ext_mark_initialized(ex);
3734 
3735 	/* note: ext4_ext_correct_indexes() isn't needed here because
3736 	 * borders are not changed
3737 	 */
3738 	ext4_ext_try_to_merge(handle, inode, path, ex);
3739 
3740 	/* Mark modified extent as dirty */
3741 	err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3742 out:
3743 	ext4_ext_show_leaf(inode, path);
3744 	return err;
3745 }
3746 
3747 static void unmap_underlying_metadata_blocks(struct block_device *bdev,
3748 			sector_t block, int count)
3749 {
3750 	int i;
3751 	for (i = 0; i < count; i++)
3752                 unmap_underlying_metadata(bdev, block + i);
3753 }
3754 
3755 /*
3756  * Handle EOFBLOCKS_FL flag, clearing it if necessary
3757  */
3758 static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
3759 			      ext4_lblk_t lblk,
3760 			      struct ext4_ext_path *path,
3761 			      unsigned int len)
3762 {
3763 	int i, depth;
3764 	struct ext4_extent_header *eh;
3765 	struct ext4_extent *last_ex;
3766 
3767 	if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3768 		return 0;
3769 
3770 	depth = ext_depth(inode);
3771 	eh = path[depth].p_hdr;
3772 
3773 	/*
3774 	 * We're going to remove EOFBLOCKS_FL entirely in future so we
3775 	 * do not care for this case anymore. Simply remove the flag
3776 	 * if there are no extents.
3777 	 */
3778 	if (unlikely(!eh->eh_entries))
3779 		goto out;
3780 	last_ex = EXT_LAST_EXTENT(eh);
3781 	/*
3782 	 * We should clear the EOFBLOCKS_FL flag if we are writing the
3783 	 * last block in the last extent in the file.  We test this by
3784 	 * first checking to see if the caller to
3785 	 * ext4_ext_get_blocks() was interested in the last block (or
3786 	 * a block beyond the last block) in the current extent.  If
3787 	 * this turns out to be false, we can bail out from this
3788 	 * function immediately.
3789 	 */
3790 	if (lblk + len < le32_to_cpu(last_ex->ee_block) +
3791 	    ext4_ext_get_actual_len(last_ex))
3792 		return 0;
3793 	/*
3794 	 * If the caller does appear to be planning to write at or
3795 	 * beyond the end of the current extent, we then test to see
3796 	 * if the current extent is the last extent in the file, by
3797 	 * checking to make sure it was reached via the rightmost node
3798 	 * at each level of the tree.
3799 	 */
3800 	for (i = depth-1; i >= 0; i--)
3801 		if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
3802 			return 0;
3803 out:
3804 	ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3805 	return ext4_mark_inode_dirty(handle, inode);
3806 }
3807 
3808 /**
3809  * ext4_find_delalloc_range: find delayed allocated block in the given range.
3810  *
3811  * Return 1 if there is a delalloc block in the range, otherwise 0.
3812  */
3813 int ext4_find_delalloc_range(struct inode *inode,
3814 			     ext4_lblk_t lblk_start,
3815 			     ext4_lblk_t lblk_end)
3816 {
3817 	struct extent_status es;
3818 
3819 	ext4_es_find_delayed_extent_range(inode, lblk_start, lblk_end, &es);
3820 	if (es.es_len == 0)
3821 		return 0; /* there is no delay extent in this tree */
3822 	else if (es.es_lblk <= lblk_start &&
3823 		 lblk_start < es.es_lblk + es.es_len)
3824 		return 1;
3825 	else if (lblk_start <= es.es_lblk && es.es_lblk <= lblk_end)
3826 		return 1;
3827 	else
3828 		return 0;
3829 }
3830 
3831 int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk)
3832 {
3833 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3834 	ext4_lblk_t lblk_start, lblk_end;
3835 	lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
3836 	lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
3837 
3838 	return ext4_find_delalloc_range(inode, lblk_start, lblk_end);
3839 }
3840 
3841 /**
3842  * Determines how many complete clusters (out of those specified by the 'map')
3843  * are under delalloc and were reserved quota for.
3844  * This function is called when we are writing out the blocks that were
3845  * originally written with their allocation delayed, but then the space was
3846  * allocated using fallocate() before the delayed allocation could be resolved.
3847  * The cases to look for are:
3848  * ('=' indicated delayed allocated blocks
3849  *  '-' indicates non-delayed allocated blocks)
3850  * (a) partial clusters towards beginning and/or end outside of allocated range
3851  *     are not delalloc'ed.
3852  *	Ex:
3853  *	|----c---=|====c====|====c====|===-c----|
3854  *	         |++++++ allocated ++++++|
3855  *	==> 4 complete clusters in above example
3856  *
3857  * (b) partial cluster (outside of allocated range) towards either end is
3858  *     marked for delayed allocation. In this case, we will exclude that
3859  *     cluster.
3860  *	Ex:
3861  *	|----====c========|========c========|
3862  *	     |++++++ allocated ++++++|
3863  *	==> 1 complete clusters in above example
3864  *
3865  *	Ex:
3866  *	|================c================|
3867  *            |++++++ allocated ++++++|
3868  *	==> 0 complete clusters in above example
3869  *
3870  * The ext4_da_update_reserve_space will be called only if we
3871  * determine here that there were some "entire" clusters that span
3872  * this 'allocated' range.
3873  * In the non-bigalloc case, this function will just end up returning num_blks
3874  * without ever calling ext4_find_delalloc_range.
3875  */
3876 static unsigned int
3877 get_reserved_cluster_alloc(struct inode *inode, ext4_lblk_t lblk_start,
3878 			   unsigned int num_blks)
3879 {
3880 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3881 	ext4_lblk_t alloc_cluster_start, alloc_cluster_end;
3882 	ext4_lblk_t lblk_from, lblk_to, c_offset;
3883 	unsigned int allocated_clusters = 0;
3884 
3885 	alloc_cluster_start = EXT4_B2C(sbi, lblk_start);
3886 	alloc_cluster_end = EXT4_B2C(sbi, lblk_start + num_blks - 1);
3887 
3888 	/* max possible clusters for this allocation */
3889 	allocated_clusters = alloc_cluster_end - alloc_cluster_start + 1;
3890 
3891 	trace_ext4_get_reserved_cluster_alloc(inode, lblk_start, num_blks);
3892 
3893 	/* Check towards left side */
3894 	c_offset = EXT4_LBLK_COFF(sbi, lblk_start);
3895 	if (c_offset) {
3896 		lblk_from = EXT4_LBLK_CMASK(sbi, lblk_start);
3897 		lblk_to = lblk_from + c_offset - 1;
3898 
3899 		if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3900 			allocated_clusters--;
3901 	}
3902 
3903 	/* Now check towards right. */
3904 	c_offset = EXT4_LBLK_COFF(sbi, lblk_start + num_blks);
3905 	if (allocated_clusters && c_offset) {
3906 		lblk_from = lblk_start + num_blks;
3907 		lblk_to = lblk_from + (sbi->s_cluster_ratio - c_offset) - 1;
3908 
3909 		if (ext4_find_delalloc_range(inode, lblk_from, lblk_to))
3910 			allocated_clusters--;
3911 	}
3912 
3913 	return allocated_clusters;
3914 }
3915 
3916 static int
3917 convert_initialized_extent(handle_t *handle, struct inode *inode,
3918 			   struct ext4_map_blocks *map,
3919 			   struct ext4_ext_path **ppath, int flags,
3920 			   unsigned int allocated, ext4_fsblk_t newblock)
3921 {
3922 	struct ext4_ext_path *path = *ppath;
3923 	struct ext4_extent *ex;
3924 	ext4_lblk_t ee_block;
3925 	unsigned int ee_len;
3926 	int depth;
3927 	int err = 0;
3928 
3929 	/*
3930 	 * Make sure that the extent is no bigger than we support with
3931 	 * unwritten extent
3932 	 */
3933 	if (map->m_len > EXT_UNWRITTEN_MAX_LEN)
3934 		map->m_len = EXT_UNWRITTEN_MAX_LEN / 2;
3935 
3936 	depth = ext_depth(inode);
3937 	ex = path[depth].p_ext;
3938 	ee_block = le32_to_cpu(ex->ee_block);
3939 	ee_len = ext4_ext_get_actual_len(ex);
3940 
3941 	ext_debug("%s: inode %lu, logical"
3942 		"block %llu, max_blocks %u\n", __func__, inode->i_ino,
3943 		  (unsigned long long)ee_block, ee_len);
3944 
3945 	if (ee_block != map->m_lblk || ee_len > map->m_len) {
3946 		err = ext4_split_convert_extents(handle, inode, map, ppath,
3947 				EXT4_GET_BLOCKS_CONVERT_UNWRITTEN);
3948 		if (err < 0)
3949 			return err;
3950 		path = ext4_find_extent(inode, map->m_lblk, ppath, 0);
3951 		if (IS_ERR(path))
3952 			return PTR_ERR(path);
3953 		depth = ext_depth(inode);
3954 		ex = path[depth].p_ext;
3955 		if (!ex) {
3956 			EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
3957 					 (unsigned long) map->m_lblk);
3958 			return -EIO;
3959 		}
3960 	}
3961 
3962 	err = ext4_ext_get_access(handle, inode, path + depth);
3963 	if (err)
3964 		return err;
3965 	/* first mark the extent as unwritten */
3966 	ext4_ext_mark_unwritten(ex);
3967 
3968 	/* note: ext4_ext_correct_indexes() isn't needed here because
3969 	 * borders are not changed
3970 	 */
3971 	ext4_ext_try_to_merge(handle, inode, path, ex);
3972 
3973 	/* Mark modified extent as dirty */
3974 	err = ext4_ext_dirty(handle, inode, path + path->p_depth);
3975 	if (err)
3976 		return err;
3977 	ext4_ext_show_leaf(inode, path);
3978 
3979 	ext4_update_inode_fsync_trans(handle, inode, 1);
3980 	err = check_eofblocks_fl(handle, inode, map->m_lblk, path, map->m_len);
3981 	if (err)
3982 		return err;
3983 	map->m_flags |= EXT4_MAP_UNWRITTEN;
3984 	if (allocated > map->m_len)
3985 		allocated = map->m_len;
3986 	map->m_len = allocated;
3987 	return allocated;
3988 }
3989 
3990 static int
3991 ext4_ext_handle_unwritten_extents(handle_t *handle, struct inode *inode,
3992 			struct ext4_map_blocks *map,
3993 			struct ext4_ext_path **ppath, int flags,
3994 			unsigned int allocated, ext4_fsblk_t newblock)
3995 {
3996 	struct ext4_ext_path *path = *ppath;
3997 	int ret = 0;
3998 	int err = 0;
3999 	ext4_io_end_t *io = ext4_inode_aio(inode);
4000 
4001 	ext_debug("ext4_ext_handle_unwritten_extents: inode %lu, logical "
4002 		  "block %llu, max_blocks %u, flags %x, allocated %u\n",
4003 		  inode->i_ino, (unsigned long long)map->m_lblk, map->m_len,
4004 		  flags, allocated);
4005 	ext4_ext_show_leaf(inode, path);
4006 
4007 	/*
4008 	 * When writing into unwritten space, we should not fail to
4009 	 * allocate metadata blocks for the new extent block if needed.
4010 	 */
4011 	flags |= EXT4_GET_BLOCKS_METADATA_NOFAIL;
4012 
4013 	trace_ext4_ext_handle_unwritten_extents(inode, map, flags,
4014 						    allocated, newblock);
4015 
4016 	/* get_block() before submit the IO, split the extent */
4017 	if (flags & EXT4_GET_BLOCKS_PRE_IO) {
4018 		ret = ext4_split_convert_extents(handle, inode, map, ppath,
4019 					 flags | EXT4_GET_BLOCKS_CONVERT);
4020 		if (ret <= 0)
4021 			goto out;
4022 		/*
4023 		 * Flag the inode(non aio case) or end_io struct (aio case)
4024 		 * that this IO needs to conversion to written when IO is
4025 		 * completed
4026 		 */
4027 		if (io)
4028 			ext4_set_io_unwritten_flag(inode, io);
4029 		else
4030 			ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
4031 		map->m_flags |= EXT4_MAP_UNWRITTEN;
4032 		goto out;
4033 	}
4034 	/* IO end_io complete, convert the filled extent to written */
4035 	if (flags & EXT4_GET_BLOCKS_CONVERT) {
4036 		ret = ext4_convert_unwritten_extents_endio(handle, inode, map,
4037 							   ppath);
4038 		if (ret >= 0) {
4039 			ext4_update_inode_fsync_trans(handle, inode, 1);
4040 			err = check_eofblocks_fl(handle, inode, map->m_lblk,
4041 						 path, map->m_len);
4042 		} else
4043 			err = ret;
4044 		map->m_flags |= EXT4_MAP_MAPPED;
4045 		map->m_pblk = newblock;
4046 		if (allocated > map->m_len)
4047 			allocated = map->m_len;
4048 		map->m_len = allocated;
4049 		goto out2;
4050 	}
4051 	/* buffered IO case */
4052 	/*
4053 	 * repeat fallocate creation request
4054 	 * we already have an unwritten extent
4055 	 */
4056 	if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT) {
4057 		map->m_flags |= EXT4_MAP_UNWRITTEN;
4058 		goto map_out;
4059 	}
4060 
4061 	/* buffered READ or buffered write_begin() lookup */
4062 	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4063 		/*
4064 		 * We have blocks reserved already.  We
4065 		 * return allocated blocks so that delalloc
4066 		 * won't do block reservation for us.  But
4067 		 * the buffer head will be unmapped so that
4068 		 * a read from the block returns 0s.
4069 		 */
4070 		map->m_flags |= EXT4_MAP_UNWRITTEN;
4071 		goto out1;
4072 	}
4073 
4074 	/* buffered write, writepage time, convert*/
4075 	ret = ext4_ext_convert_to_initialized(handle, inode, map, ppath, flags);
4076 	if (ret >= 0)
4077 		ext4_update_inode_fsync_trans(handle, inode, 1);
4078 out:
4079 	if (ret <= 0) {
4080 		err = ret;
4081 		goto out2;
4082 	} else
4083 		allocated = ret;
4084 	map->m_flags |= EXT4_MAP_NEW;
4085 	/*
4086 	 * if we allocated more blocks than requested
4087 	 * we need to make sure we unmap the extra block
4088 	 * allocated. The actual needed block will get
4089 	 * unmapped later when we find the buffer_head marked
4090 	 * new.
4091 	 */
4092 	if (allocated > map->m_len) {
4093 		unmap_underlying_metadata_blocks(inode->i_sb->s_bdev,
4094 					newblock + map->m_len,
4095 					allocated - map->m_len);
4096 		allocated = map->m_len;
4097 	}
4098 	map->m_len = allocated;
4099 
4100 	/*
4101 	 * If we have done fallocate with the offset that is already
4102 	 * delayed allocated, we would have block reservation
4103 	 * and quota reservation done in the delayed write path.
4104 	 * But fallocate would have already updated quota and block
4105 	 * count for this offset. So cancel these reservation
4106 	 */
4107 	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4108 		unsigned int reserved_clusters;
4109 		reserved_clusters = get_reserved_cluster_alloc(inode,
4110 				map->m_lblk, map->m_len);
4111 		if (reserved_clusters)
4112 			ext4_da_update_reserve_space(inode,
4113 						     reserved_clusters,
4114 						     0);
4115 	}
4116 
4117 map_out:
4118 	map->m_flags |= EXT4_MAP_MAPPED;
4119 	if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0) {
4120 		err = check_eofblocks_fl(handle, inode, map->m_lblk, path,
4121 					 map->m_len);
4122 		if (err < 0)
4123 			goto out2;
4124 	}
4125 out1:
4126 	if (allocated > map->m_len)
4127 		allocated = map->m_len;
4128 	ext4_ext_show_leaf(inode, path);
4129 	map->m_pblk = newblock;
4130 	map->m_len = allocated;
4131 out2:
4132 	return err ? err : allocated;
4133 }
4134 
4135 /*
4136  * get_implied_cluster_alloc - check to see if the requested
4137  * allocation (in the map structure) overlaps with a cluster already
4138  * allocated in an extent.
4139  *	@sb	The filesystem superblock structure
4140  *	@map	The requested lblk->pblk mapping
4141  *	@ex	The extent structure which might contain an implied
4142  *			cluster allocation
4143  *
4144  * This function is called by ext4_ext_map_blocks() after we failed to
4145  * find blocks that were already in the inode's extent tree.  Hence,
4146  * we know that the beginning of the requested region cannot overlap
4147  * the extent from the inode's extent tree.  There are three cases we
4148  * want to catch.  The first is this case:
4149  *
4150  *		 |--- cluster # N--|
4151  *    |--- extent ---|	|---- requested region ---|
4152  *			|==========|
4153  *
4154  * The second case that we need to test for is this one:
4155  *
4156  *   |--------- cluster # N ----------------|
4157  *	   |--- requested region --|   |------- extent ----|
4158  *	   |=======================|
4159  *
4160  * The third case is when the requested region lies between two extents
4161  * within the same cluster:
4162  *          |------------- cluster # N-------------|
4163  * |----- ex -----|                  |---- ex_right ----|
4164  *                  |------ requested region ------|
4165  *                  |================|
4166  *
4167  * In each of the above cases, we need to set the map->m_pblk and
4168  * map->m_len so it corresponds to the return the extent labelled as
4169  * "|====|" from cluster #N, since it is already in use for data in
4170  * cluster EXT4_B2C(sbi, map->m_lblk).	We will then return 1 to
4171  * signal to ext4_ext_map_blocks() that map->m_pblk should be treated
4172  * as a new "allocated" block region.  Otherwise, we will return 0 and
4173  * ext4_ext_map_blocks() will then allocate one or more new clusters
4174  * by calling ext4_mb_new_blocks().
4175  */
4176 static int get_implied_cluster_alloc(struct super_block *sb,
4177 				     struct ext4_map_blocks *map,
4178 				     struct ext4_extent *ex,
4179 				     struct ext4_ext_path *path)
4180 {
4181 	struct ext4_sb_info *sbi = EXT4_SB(sb);
4182 	ext4_lblk_t c_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4183 	ext4_lblk_t ex_cluster_start, ex_cluster_end;
4184 	ext4_lblk_t rr_cluster_start;
4185 	ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4186 	ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4187 	unsigned short ee_len = ext4_ext_get_actual_len(ex);
4188 
4189 	/* The extent passed in that we are trying to match */
4190 	ex_cluster_start = EXT4_B2C(sbi, ee_block);
4191 	ex_cluster_end = EXT4_B2C(sbi, ee_block + ee_len - 1);
4192 
4193 	/* The requested region passed into ext4_map_blocks() */
4194 	rr_cluster_start = EXT4_B2C(sbi, map->m_lblk);
4195 
4196 	if ((rr_cluster_start == ex_cluster_end) ||
4197 	    (rr_cluster_start == ex_cluster_start)) {
4198 		if (rr_cluster_start == ex_cluster_end)
4199 			ee_start += ee_len - 1;
4200 		map->m_pblk = EXT4_PBLK_CMASK(sbi, ee_start) + c_offset;
4201 		map->m_len = min(map->m_len,
4202 				 (unsigned) sbi->s_cluster_ratio - c_offset);
4203 		/*
4204 		 * Check for and handle this case:
4205 		 *
4206 		 *   |--------- cluster # N-------------|
4207 		 *		       |------- extent ----|
4208 		 *	   |--- requested region ---|
4209 		 *	   |===========|
4210 		 */
4211 
4212 		if (map->m_lblk < ee_block)
4213 			map->m_len = min(map->m_len, ee_block - map->m_lblk);
4214 
4215 		/*
4216 		 * Check for the case where there is already another allocated
4217 		 * block to the right of 'ex' but before the end of the cluster.
4218 		 *
4219 		 *          |------------- cluster # N-------------|
4220 		 * |----- ex -----|                  |---- ex_right ----|
4221 		 *                  |------ requested region ------|
4222 		 *                  |================|
4223 		 */
4224 		if (map->m_lblk > ee_block) {
4225 			ext4_lblk_t next = ext4_ext_next_allocated_block(path);
4226 			map->m_len = min(map->m_len, next - map->m_lblk);
4227 		}
4228 
4229 		trace_ext4_get_implied_cluster_alloc_exit(sb, map, 1);
4230 		return 1;
4231 	}
4232 
4233 	trace_ext4_get_implied_cluster_alloc_exit(sb, map, 0);
4234 	return 0;
4235 }
4236 
4237 
4238 /*
4239  * Block allocation/map/preallocation routine for extents based files
4240  *
4241  *
4242  * Need to be called with
4243  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system block
4244  * (ie, create is zero). Otherwise down_write(&EXT4_I(inode)->i_data_sem)
4245  *
4246  * return > 0, number of of blocks already mapped/allocated
4247  *          if create == 0 and these are pre-allocated blocks
4248  *          	buffer head is unmapped
4249  *          otherwise blocks are mapped
4250  *
4251  * return = 0, if plain look up failed (blocks have not been allocated)
4252  *          buffer head is unmapped
4253  *
4254  * return < 0, error case.
4255  */
4256 int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
4257 			struct ext4_map_blocks *map, int flags)
4258 {
4259 	struct ext4_ext_path *path = NULL;
4260 	struct ext4_extent newex, *ex, *ex2;
4261 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4262 	ext4_fsblk_t newblock = 0;
4263 	int free_on_err = 0, err = 0, depth, ret;
4264 	unsigned int allocated = 0, offset = 0;
4265 	unsigned int allocated_clusters = 0;
4266 	struct ext4_allocation_request ar;
4267 	ext4_io_end_t *io = ext4_inode_aio(inode);
4268 	ext4_lblk_t cluster_offset;
4269 	int set_unwritten = 0;
4270 
4271 	ext_debug("blocks %u/%u requested for inode %lu\n",
4272 		  map->m_lblk, map->m_len, inode->i_ino);
4273 	trace_ext4_ext_map_blocks_enter(inode, map->m_lblk, map->m_len, flags);
4274 
4275 	/* find extent for this block */
4276 	path = ext4_find_extent(inode, map->m_lblk, NULL, 0);
4277 	if (IS_ERR(path)) {
4278 		err = PTR_ERR(path);
4279 		path = NULL;
4280 		goto out2;
4281 	}
4282 
4283 	depth = ext_depth(inode);
4284 
4285 	/*
4286 	 * consistent leaf must not be empty;
4287 	 * this situation is possible, though, _during_ tree modification;
4288 	 * this is why assert can't be put in ext4_find_extent()
4289 	 */
4290 	if (unlikely(path[depth].p_ext == NULL && depth != 0)) {
4291 		EXT4_ERROR_INODE(inode, "bad extent address "
4292 				 "lblock: %lu, depth: %d pblock %lld",
4293 				 (unsigned long) map->m_lblk, depth,
4294 				 path[depth].p_block);
4295 		err = -EIO;
4296 		goto out2;
4297 	}
4298 
4299 	ex = path[depth].p_ext;
4300 	if (ex) {
4301 		ext4_lblk_t ee_block = le32_to_cpu(ex->ee_block);
4302 		ext4_fsblk_t ee_start = ext4_ext_pblock(ex);
4303 		unsigned short ee_len;
4304 
4305 
4306 		/*
4307 		 * unwritten extents are treated as holes, except that
4308 		 * we split out initialized portions during a write.
4309 		 */
4310 		ee_len = ext4_ext_get_actual_len(ex);
4311 
4312 		trace_ext4_ext_show_extent(inode, ee_block, ee_start, ee_len);
4313 
4314 		/* if found extent covers block, simply return it */
4315 		if (in_range(map->m_lblk, ee_block, ee_len)) {
4316 			newblock = map->m_lblk - ee_block + ee_start;
4317 			/* number of remaining blocks in the extent */
4318 			allocated = ee_len - (map->m_lblk - ee_block);
4319 			ext_debug("%u fit into %u:%d -> %llu\n", map->m_lblk,
4320 				  ee_block, ee_len, newblock);
4321 
4322 			/*
4323 			 * If the extent is initialized check whether the
4324 			 * caller wants to convert it to unwritten.
4325 			 */
4326 			if ((!ext4_ext_is_unwritten(ex)) &&
4327 			    (flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN)) {
4328 				allocated = convert_initialized_extent(
4329 						handle, inode, map, &path,
4330 						flags, allocated, newblock);
4331 				goto out2;
4332 			} else if (!ext4_ext_is_unwritten(ex))
4333 				goto out;
4334 
4335 			ret = ext4_ext_handle_unwritten_extents(
4336 				handle, inode, map, &path, flags,
4337 				allocated, newblock);
4338 			if (ret < 0)
4339 				err = ret;
4340 			else
4341 				allocated = ret;
4342 			goto out2;
4343 		}
4344 	}
4345 
4346 	if ((sbi->s_cluster_ratio > 1) &&
4347 	    ext4_find_delalloc_cluster(inode, map->m_lblk))
4348 		map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4349 
4350 	/*
4351 	 * requested block isn't allocated yet;
4352 	 * we couldn't try to create block if create flag is zero
4353 	 */
4354 	if ((flags & EXT4_GET_BLOCKS_CREATE) == 0) {
4355 		/*
4356 		 * put just found gap into cache to speed up
4357 		 * subsequent requests
4358 		 */
4359 		if ((flags & EXT4_GET_BLOCKS_NO_PUT_HOLE) == 0)
4360 			ext4_ext_put_gap_in_cache(inode, path, map->m_lblk);
4361 		goto out2;
4362 	}
4363 
4364 	/*
4365 	 * Okay, we need to do block allocation.
4366 	 */
4367 	map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
4368 	newex.ee_block = cpu_to_le32(map->m_lblk);
4369 	cluster_offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4370 
4371 	/*
4372 	 * If we are doing bigalloc, check to see if the extent returned
4373 	 * by ext4_find_extent() implies a cluster we can use.
4374 	 */
4375 	if (cluster_offset && ex &&
4376 	    get_implied_cluster_alloc(inode->i_sb, map, ex, path)) {
4377 		ar.len = allocated = map->m_len;
4378 		newblock = map->m_pblk;
4379 		map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4380 		goto got_allocated_blocks;
4381 	}
4382 
4383 	/* find neighbour allocated blocks */
4384 	ar.lleft = map->m_lblk;
4385 	err = ext4_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
4386 	if (err)
4387 		goto out2;
4388 	ar.lright = map->m_lblk;
4389 	ex2 = NULL;
4390 	err = ext4_ext_search_right(inode, path, &ar.lright, &ar.pright, &ex2);
4391 	if (err)
4392 		goto out2;
4393 
4394 	/* Check if the extent after searching to the right implies a
4395 	 * cluster we can use. */
4396 	if ((sbi->s_cluster_ratio > 1) && ex2 &&
4397 	    get_implied_cluster_alloc(inode->i_sb, map, ex2, path)) {
4398 		ar.len = allocated = map->m_len;
4399 		newblock = map->m_pblk;
4400 		map->m_flags |= EXT4_MAP_FROM_CLUSTER;
4401 		goto got_allocated_blocks;
4402 	}
4403 
4404 	/*
4405 	 * See if request is beyond maximum number of blocks we can have in
4406 	 * a single extent. For an initialized extent this limit is
4407 	 * EXT_INIT_MAX_LEN and for an unwritten extent this limit is
4408 	 * EXT_UNWRITTEN_MAX_LEN.
4409 	 */
4410 	if (map->m_len > EXT_INIT_MAX_LEN &&
4411 	    !(flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4412 		map->m_len = EXT_INIT_MAX_LEN;
4413 	else if (map->m_len > EXT_UNWRITTEN_MAX_LEN &&
4414 		 (flags & EXT4_GET_BLOCKS_UNWRIT_EXT))
4415 		map->m_len = EXT_UNWRITTEN_MAX_LEN;
4416 
4417 	/* Check if we can really insert (m_lblk)::(m_lblk + m_len) extent */
4418 	newex.ee_len = cpu_to_le16(map->m_len);
4419 	err = ext4_ext_check_overlap(sbi, inode, &newex, path);
4420 	if (err)
4421 		allocated = ext4_ext_get_actual_len(&newex);
4422 	else
4423 		allocated = map->m_len;
4424 
4425 	/* allocate new block */
4426 	ar.inode = inode;
4427 	ar.goal = ext4_ext_find_goal(inode, path, map->m_lblk);
4428 	ar.logical = map->m_lblk;
4429 	/*
4430 	 * We calculate the offset from the beginning of the cluster
4431 	 * for the logical block number, since when we allocate a
4432 	 * physical cluster, the physical block should start at the
4433 	 * same offset from the beginning of the cluster.  This is
4434 	 * needed so that future calls to get_implied_cluster_alloc()
4435 	 * work correctly.
4436 	 */
4437 	offset = EXT4_LBLK_COFF(sbi, map->m_lblk);
4438 	ar.len = EXT4_NUM_B2C(sbi, offset+allocated);
4439 	ar.goal -= offset;
4440 	ar.logical -= offset;
4441 	if (S_ISREG(inode->i_mode))
4442 		ar.flags = EXT4_MB_HINT_DATA;
4443 	else
4444 		/* disable in-core preallocation for non-regular files */
4445 		ar.flags = 0;
4446 	if (flags & EXT4_GET_BLOCKS_NO_NORMALIZE)
4447 		ar.flags |= EXT4_MB_HINT_NOPREALLOC;
4448 	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
4449 		ar.flags |= EXT4_MB_DELALLOC_RESERVED;
4450 	newblock = ext4_mb_new_blocks(handle, &ar, &err);
4451 	if (!newblock)
4452 		goto out2;
4453 	ext_debug("allocate new block: goal %llu, found %llu/%u\n",
4454 		  ar.goal, newblock, allocated);
4455 	free_on_err = 1;
4456 	allocated_clusters = ar.len;
4457 	ar.len = EXT4_C2B(sbi, ar.len) - offset;
4458 	if (ar.len > allocated)
4459 		ar.len = allocated;
4460 
4461 got_allocated_blocks:
4462 	/* try to insert new extent into found leaf and return */
4463 	ext4_ext_store_pblock(&newex, newblock + offset);
4464 	newex.ee_len = cpu_to_le16(ar.len);
4465 	/* Mark unwritten */
4466 	if (flags & EXT4_GET_BLOCKS_UNWRIT_EXT){
4467 		ext4_ext_mark_unwritten(&newex);
4468 		map->m_flags |= EXT4_MAP_UNWRITTEN;
4469 		/*
4470 		 * io_end structure was created for every IO write to an
4471 		 * unwritten extent. To avoid unnecessary conversion,
4472 		 * here we flag the IO that really needs the conversion.
4473 		 * For non asycn direct IO case, flag the inode state
4474 		 * that we need to perform conversion when IO is done.
4475 		 */
4476 		if (flags & EXT4_GET_BLOCKS_PRE_IO)
4477 			set_unwritten = 1;
4478 	}
4479 
4480 	err = 0;
4481 	if ((flags & EXT4_GET_BLOCKS_KEEP_SIZE) == 0)
4482 		err = check_eofblocks_fl(handle, inode, map->m_lblk,
4483 					 path, ar.len);
4484 	if (!err)
4485 		err = ext4_ext_insert_extent(handle, inode, &path,
4486 					     &newex, flags);
4487 
4488 	if (!err && set_unwritten) {
4489 		if (io)
4490 			ext4_set_io_unwritten_flag(inode, io);
4491 		else
4492 			ext4_set_inode_state(inode,
4493 					     EXT4_STATE_DIO_UNWRITTEN);
4494 	}
4495 
4496 	if (err && free_on_err) {
4497 		int fb_flags = flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE ?
4498 			EXT4_FREE_BLOCKS_NO_QUOT_UPDATE : 0;
4499 		/* free data blocks we just allocated */
4500 		/* not a good idea to call discard here directly,
4501 		 * but otherwise we'd need to call it every free() */
4502 		ext4_discard_preallocations(inode);
4503 		ext4_free_blocks(handle, inode, NULL, newblock,
4504 				 EXT4_C2B(sbi, allocated_clusters), fb_flags);
4505 		goto out2;
4506 	}
4507 
4508 	/* previous routine could use block we allocated */
4509 	newblock = ext4_ext_pblock(&newex);
4510 	allocated = ext4_ext_get_actual_len(&newex);
4511 	if (allocated > map->m_len)
4512 		allocated = map->m_len;
4513 	map->m_flags |= EXT4_MAP_NEW;
4514 
4515 	/*
4516 	 * Update reserved blocks/metadata blocks after successful
4517 	 * block allocation which had been deferred till now.
4518 	 */
4519 	if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
4520 		unsigned int reserved_clusters;
4521 		/*
4522 		 * Check how many clusters we had reserved this allocated range
4523 		 */
4524 		reserved_clusters = get_reserved_cluster_alloc(inode,
4525 						map->m_lblk, allocated);
4526 		if (map->m_flags & EXT4_MAP_FROM_CLUSTER) {
4527 			if (reserved_clusters) {
4528 				/*
4529 				 * We have clusters reserved for this range.
4530 				 * But since we are not doing actual allocation
4531 				 * and are simply using blocks from previously
4532 				 * allocated cluster, we should release the
4533 				 * reservation and not claim quota.
4534 				 */
4535 				ext4_da_update_reserve_space(inode,
4536 						reserved_clusters, 0);
4537 			}
4538 		} else {
4539 			BUG_ON(allocated_clusters < reserved_clusters);
4540 			if (reserved_clusters < allocated_clusters) {
4541 				struct ext4_inode_info *ei = EXT4_I(inode);
4542 				int reservation = allocated_clusters -
4543 						  reserved_clusters;
4544 				/*
4545 				 * It seems we claimed few clusters outside of
4546 				 * the range of this allocation. We should give
4547 				 * it back to the reservation pool. This can
4548 				 * happen in the following case:
4549 				 *
4550 				 * * Suppose s_cluster_ratio is 4 (i.e., each
4551 				 *   cluster has 4 blocks. Thus, the clusters
4552 				 *   are [0-3],[4-7],[8-11]...
4553 				 * * First comes delayed allocation write for
4554 				 *   logical blocks 10 & 11. Since there were no
4555 				 *   previous delayed allocated blocks in the
4556 				 *   range [8-11], we would reserve 1 cluster
4557 				 *   for this write.
4558 				 * * Next comes write for logical blocks 3 to 8.
4559 				 *   In this case, we will reserve 2 clusters
4560 				 *   (for [0-3] and [4-7]; and not for [8-11] as
4561 				 *   that range has a delayed allocated blocks.
4562 				 *   Thus total reserved clusters now becomes 3.
4563 				 * * Now, during the delayed allocation writeout
4564 				 *   time, we will first write blocks [3-8] and
4565 				 *   allocate 3 clusters for writing these
4566 				 *   blocks. Also, we would claim all these
4567 				 *   three clusters above.
4568 				 * * Now when we come here to writeout the
4569 				 *   blocks [10-11], we would expect to claim
4570 				 *   the reservation of 1 cluster we had made
4571 				 *   (and we would claim it since there are no
4572 				 *   more delayed allocated blocks in the range
4573 				 *   [8-11]. But our reserved cluster count had
4574 				 *   already gone to 0.
4575 				 *
4576 				 *   Thus, at the step 4 above when we determine
4577 				 *   that there are still some unwritten delayed
4578 				 *   allocated blocks outside of our current
4579 				 *   block range, we should increment the
4580 				 *   reserved clusters count so that when the
4581 				 *   remaining blocks finally gets written, we
4582 				 *   could claim them.
4583 				 */
4584 				dquot_reserve_block(inode,
4585 						EXT4_C2B(sbi, reservation));
4586 				spin_lock(&ei->i_block_reservation_lock);
4587 				ei->i_reserved_data_blocks += reservation;
4588 				spin_unlock(&ei->i_block_reservation_lock);
4589 			}
4590 			/*
4591 			 * We will claim quota for all newly allocated blocks.
4592 			 * We're updating the reserved space *after* the
4593 			 * correction above so we do not accidentally free
4594 			 * all the metadata reservation because we might
4595 			 * actually need it later on.
4596 			 */
4597 			ext4_da_update_reserve_space(inode, allocated_clusters,
4598 							1);
4599 		}
4600 	}
4601 
4602 	/*
4603 	 * Cache the extent and update transaction to commit on fdatasync only
4604 	 * when it is _not_ an unwritten extent.
4605 	 */
4606 	if ((flags & EXT4_GET_BLOCKS_UNWRIT_EXT) == 0)
4607 		ext4_update_inode_fsync_trans(handle, inode, 1);
4608 	else
4609 		ext4_update_inode_fsync_trans(handle, inode, 0);
4610 out:
4611 	if (allocated > map->m_len)
4612 		allocated = map->m_len;
4613 	ext4_ext_show_leaf(inode, path);
4614 	map->m_flags |= EXT4_MAP_MAPPED;
4615 	map->m_pblk = newblock;
4616 	map->m_len = allocated;
4617 out2:
4618 	ext4_ext_drop_refs(path);
4619 	kfree(path);
4620 
4621 	trace_ext4_ext_map_blocks_exit(inode, flags, map,
4622 				       err ? err : allocated);
4623 	ext4_es_lru_add(inode);
4624 	return err ? err : allocated;
4625 }
4626 
4627 void ext4_ext_truncate(handle_t *handle, struct inode *inode)
4628 {
4629 	struct super_block *sb = inode->i_sb;
4630 	ext4_lblk_t last_block;
4631 	int err = 0;
4632 
4633 	/*
4634 	 * TODO: optimization is possible here.
4635 	 * Probably we need not scan at all,
4636 	 * because page truncation is enough.
4637 	 */
4638 
4639 	/* we have to know where to truncate from in crash case */
4640 	EXT4_I(inode)->i_disksize = inode->i_size;
4641 	ext4_mark_inode_dirty(handle, inode);
4642 
4643 	last_block = (inode->i_size + sb->s_blocksize - 1)
4644 			>> EXT4_BLOCK_SIZE_BITS(sb);
4645 retry:
4646 	err = ext4_es_remove_extent(inode, last_block,
4647 				    EXT_MAX_BLOCKS - last_block);
4648 	if (err == -ENOMEM) {
4649 		cond_resched();
4650 		congestion_wait(BLK_RW_ASYNC, HZ/50);
4651 		goto retry;
4652 	}
4653 	if (err) {
4654 		ext4_std_error(inode->i_sb, err);
4655 		return;
4656 	}
4657 	err = ext4_ext_remove_space(inode, last_block, EXT_MAX_BLOCKS - 1);
4658 	ext4_std_error(inode->i_sb, err);
4659 }
4660 
4661 static int ext4_alloc_file_blocks(struct file *file, ext4_lblk_t offset,
4662 				  ext4_lblk_t len, loff_t new_size,
4663 				  int flags, int mode)
4664 {
4665 	struct inode *inode = file_inode(file);
4666 	handle_t *handle;
4667 	int ret = 0;
4668 	int ret2 = 0;
4669 	int retries = 0;
4670 	struct ext4_map_blocks map;
4671 	unsigned int credits;
4672 	loff_t epos;
4673 
4674 	map.m_lblk = offset;
4675 	map.m_len = len;
4676 	/*
4677 	 * Don't normalize the request if it can fit in one extent so
4678 	 * that it doesn't get unnecessarily split into multiple
4679 	 * extents.
4680 	 */
4681 	if (len <= EXT_UNWRITTEN_MAX_LEN)
4682 		flags |= EXT4_GET_BLOCKS_NO_NORMALIZE;
4683 
4684 	/*
4685 	 * credits to insert 1 extent into extent tree
4686 	 */
4687 	credits = ext4_chunk_trans_blocks(inode, len);
4688 
4689 retry:
4690 	while (ret >= 0 && len) {
4691 		handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
4692 					    credits);
4693 		if (IS_ERR(handle)) {
4694 			ret = PTR_ERR(handle);
4695 			break;
4696 		}
4697 		ret = ext4_map_blocks(handle, inode, &map, flags);
4698 		if (ret <= 0) {
4699 			ext4_debug("inode #%lu: block %u: len %u: "
4700 				   "ext4_ext_map_blocks returned %d",
4701 				   inode->i_ino, map.m_lblk,
4702 				   map.m_len, ret);
4703 			ext4_mark_inode_dirty(handle, inode);
4704 			ret2 = ext4_journal_stop(handle);
4705 			break;
4706 		}
4707 		map.m_lblk += ret;
4708 		map.m_len = len = len - ret;
4709 		epos = (loff_t)map.m_lblk << inode->i_blkbits;
4710 		inode->i_ctime = ext4_current_time(inode);
4711 		if (new_size) {
4712 			if (epos > new_size)
4713 				epos = new_size;
4714 			if (ext4_update_inode_size(inode, epos) & 0x1)
4715 				inode->i_mtime = inode->i_ctime;
4716 		} else {
4717 			if (epos > inode->i_size)
4718 				ext4_set_inode_flag(inode,
4719 						    EXT4_INODE_EOFBLOCKS);
4720 		}
4721 		ext4_mark_inode_dirty(handle, inode);
4722 		ret2 = ext4_journal_stop(handle);
4723 		if (ret2)
4724 			break;
4725 	}
4726 	if (ret == -ENOSPC &&
4727 			ext4_should_retry_alloc(inode->i_sb, &retries)) {
4728 		ret = 0;
4729 		goto retry;
4730 	}
4731 
4732 	return ret > 0 ? ret2 : ret;
4733 }
4734 
4735 static long ext4_zero_range(struct file *file, loff_t offset,
4736 			    loff_t len, int mode)
4737 {
4738 	struct inode *inode = file_inode(file);
4739 	handle_t *handle = NULL;
4740 	unsigned int max_blocks;
4741 	loff_t new_size = 0;
4742 	int ret = 0;
4743 	int flags;
4744 	int credits;
4745 	int partial_begin, partial_end;
4746 	loff_t start, end;
4747 	ext4_lblk_t lblk;
4748 	struct address_space *mapping = inode->i_mapping;
4749 	unsigned int blkbits = inode->i_blkbits;
4750 
4751 	trace_ext4_zero_range(inode, offset, len, mode);
4752 
4753 	if (!S_ISREG(inode->i_mode))
4754 		return -EINVAL;
4755 
4756 	/* Call ext4_force_commit to flush all data in case of data=journal. */
4757 	if (ext4_should_journal_data(inode)) {
4758 		ret = ext4_force_commit(inode->i_sb);
4759 		if (ret)
4760 			return ret;
4761 	}
4762 
4763 	/*
4764 	 * Write out all dirty pages to avoid race conditions
4765 	 * Then release them.
4766 	 */
4767 	if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
4768 		ret = filemap_write_and_wait_range(mapping, offset,
4769 						   offset + len - 1);
4770 		if (ret)
4771 			return ret;
4772 	}
4773 
4774 	/*
4775 	 * Round up offset. This is not fallocate, we neet to zero out
4776 	 * blocks, so convert interior block aligned part of the range to
4777 	 * unwritten and possibly manually zero out unaligned parts of the
4778 	 * range.
4779 	 */
4780 	start = round_up(offset, 1 << blkbits);
4781 	end = round_down((offset + len), 1 << blkbits);
4782 
4783 	if (start < offset || end > offset + len)
4784 		return -EINVAL;
4785 	partial_begin = offset & ((1 << blkbits) - 1);
4786 	partial_end = (offset + len) & ((1 << blkbits) - 1);
4787 
4788 	lblk = start >> blkbits;
4789 	max_blocks = (end >> blkbits);
4790 	if (max_blocks < lblk)
4791 		max_blocks = 0;
4792 	else
4793 		max_blocks -= lblk;
4794 
4795 	flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT |
4796 		EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
4797 		EXT4_EX_NOCACHE;
4798 	if (mode & FALLOC_FL_KEEP_SIZE)
4799 		flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4800 
4801 	mutex_lock(&inode->i_mutex);
4802 
4803 	/*
4804 	 * Indirect files do not support unwritten extnets
4805 	 */
4806 	if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4807 		ret = -EOPNOTSUPP;
4808 		goto out_mutex;
4809 	}
4810 
4811 	if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4812 	     offset + len > i_size_read(inode)) {
4813 		new_size = offset + len;
4814 		ret = inode_newsize_ok(inode, new_size);
4815 		if (ret)
4816 			goto out_mutex;
4817 		/*
4818 		 * If we have a partial block after EOF we have to allocate
4819 		 * the entire block.
4820 		 */
4821 		if (partial_end)
4822 			max_blocks += 1;
4823 	}
4824 
4825 	if (max_blocks > 0) {
4826 
4827 		/* Now release the pages and zero block aligned part of pages*/
4828 		truncate_pagecache_range(inode, start, end - 1);
4829 		inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4830 
4831 		/* Wait all existing dio workers, newcomers will block on i_mutex */
4832 		ext4_inode_block_unlocked_dio(inode);
4833 		inode_dio_wait(inode);
4834 
4835 		ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4836 					     flags, mode);
4837 		if (ret)
4838 			goto out_dio;
4839 		/*
4840 		 * Remove entire range from the extent status tree.
4841 		 *
4842 		 * ext4_es_remove_extent(inode, lblk, max_blocks) is
4843 		 * NOT sufficient.  I'm not sure why this is the case,
4844 		 * but let's be conservative and remove the extent
4845 		 * status tree for the entire inode.  There should be
4846 		 * no outstanding delalloc extents thanks to the
4847 		 * filemap_write_and_wait_range() call above.
4848 		 */
4849 		ret = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
4850 		if (ret)
4851 			goto out_dio;
4852 	}
4853 	if (!partial_begin && !partial_end)
4854 		goto out_dio;
4855 
4856 	/*
4857 	 * In worst case we have to writeout two nonadjacent unwritten
4858 	 * blocks and update the inode
4859 	 */
4860 	credits = (2 * ext4_ext_index_trans_blocks(inode, 2)) + 1;
4861 	if (ext4_should_journal_data(inode))
4862 		credits += 2;
4863 	handle = ext4_journal_start(inode, EXT4_HT_MISC, credits);
4864 	if (IS_ERR(handle)) {
4865 		ret = PTR_ERR(handle);
4866 		ext4_std_error(inode->i_sb, ret);
4867 		goto out_dio;
4868 	}
4869 
4870 	inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4871 	if (new_size) {
4872 		ext4_update_inode_size(inode, new_size);
4873 	} else {
4874 		/*
4875 		* Mark that we allocate beyond EOF so the subsequent truncate
4876 		* can proceed even if the new size is the same as i_size.
4877 		*/
4878 		if ((offset + len) > i_size_read(inode))
4879 			ext4_set_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
4880 	}
4881 	ext4_mark_inode_dirty(handle, inode);
4882 
4883 	/* Zero out partial block at the edges of the range */
4884 	ret = ext4_zero_partial_blocks(handle, inode, offset, len);
4885 
4886 	if (file->f_flags & O_SYNC)
4887 		ext4_handle_sync(handle);
4888 
4889 	ext4_journal_stop(handle);
4890 out_dio:
4891 	ext4_inode_resume_unlocked_dio(inode);
4892 out_mutex:
4893 	mutex_unlock(&inode->i_mutex);
4894 	return ret;
4895 }
4896 
4897 /*
4898  * preallocate space for a file. This implements ext4's fallocate file
4899  * operation, which gets called from sys_fallocate system call.
4900  * For block-mapped files, posix_fallocate should fall back to the method
4901  * of writing zeroes to the required new blocks (the same behavior which is
4902  * expected for file systems which do not support fallocate() system call).
4903  */
4904 long ext4_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
4905 {
4906 	struct inode *inode = file_inode(file);
4907 	loff_t new_size = 0;
4908 	unsigned int max_blocks;
4909 	int ret = 0;
4910 	int flags;
4911 	ext4_lblk_t lblk;
4912 	unsigned int blkbits = inode->i_blkbits;
4913 
4914 	/* Return error if mode is not supported */
4915 	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
4916 		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
4917 		return -EOPNOTSUPP;
4918 
4919 	if (mode & FALLOC_FL_PUNCH_HOLE)
4920 		return ext4_punch_hole(inode, offset, len);
4921 
4922 	ret = ext4_convert_inline_data(inode);
4923 	if (ret)
4924 		return ret;
4925 
4926 	/*
4927 	 * currently supporting (pre)allocate mode for extent-based
4928 	 * files _only_
4929 	 */
4930 	if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4931 		return -EOPNOTSUPP;
4932 
4933 	if (mode & FALLOC_FL_COLLAPSE_RANGE)
4934 		return ext4_collapse_range(inode, offset, len);
4935 
4936 	if (mode & FALLOC_FL_ZERO_RANGE)
4937 		return ext4_zero_range(file, offset, len, mode);
4938 
4939 	trace_ext4_fallocate_enter(inode, offset, len, mode);
4940 	lblk = offset >> blkbits;
4941 	/*
4942 	 * We can't just convert len to max_blocks because
4943 	 * If blocksize = 4096 offset = 3072 and len = 2048
4944 	 */
4945 	max_blocks = (EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits)
4946 		- lblk;
4947 
4948 	flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT;
4949 	if (mode & FALLOC_FL_KEEP_SIZE)
4950 		flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
4951 
4952 	mutex_lock(&inode->i_mutex);
4953 
4954 	if (!(mode & FALLOC_FL_KEEP_SIZE) &&
4955 	     offset + len > i_size_read(inode)) {
4956 		new_size = offset + len;
4957 		ret = inode_newsize_ok(inode, new_size);
4958 		if (ret)
4959 			goto out;
4960 	}
4961 
4962 	ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
4963 				     flags, mode);
4964 	if (ret)
4965 		goto out;
4966 
4967 	if (file->f_flags & O_SYNC && EXT4_SB(inode->i_sb)->s_journal) {
4968 		ret = jbd2_complete_transaction(EXT4_SB(inode->i_sb)->s_journal,
4969 						EXT4_I(inode)->i_sync_tid);
4970 	}
4971 out:
4972 	mutex_unlock(&inode->i_mutex);
4973 	trace_ext4_fallocate_exit(inode, offset, max_blocks, ret);
4974 	return ret;
4975 }
4976 
4977 /*
4978  * This function convert a range of blocks to written extents
4979  * The caller of this function will pass the start offset and the size.
4980  * all unwritten extents within this range will be converted to
4981  * written extents.
4982  *
4983  * This function is called from the direct IO end io call back
4984  * function, to convert the fallocated extents after IO is completed.
4985  * Returns 0 on success.
4986  */
4987 int ext4_convert_unwritten_extents(handle_t *handle, struct inode *inode,
4988 				   loff_t offset, ssize_t len)
4989 {
4990 	unsigned int max_blocks;
4991 	int ret = 0;
4992 	int ret2 = 0;
4993 	struct ext4_map_blocks map;
4994 	unsigned int credits, blkbits = inode->i_blkbits;
4995 
4996 	map.m_lblk = offset >> blkbits;
4997 	/*
4998 	 * We can't just convert len to max_blocks because
4999 	 * If blocksize = 4096 offset = 3072 and len = 2048
5000 	 */
5001 	max_blocks = ((EXT4_BLOCK_ALIGN(len + offset, blkbits) >> blkbits) -
5002 		      map.m_lblk);
5003 	/*
5004 	 * This is somewhat ugly but the idea is clear: When transaction is
5005 	 * reserved, everything goes into it. Otherwise we rather start several
5006 	 * smaller transactions for conversion of each extent separately.
5007 	 */
5008 	if (handle) {
5009 		handle = ext4_journal_start_reserved(handle,
5010 						     EXT4_HT_EXT_CONVERT);
5011 		if (IS_ERR(handle))
5012 			return PTR_ERR(handle);
5013 		credits = 0;
5014 	} else {
5015 		/*
5016 		 * credits to insert 1 extent into extent tree
5017 		 */
5018 		credits = ext4_chunk_trans_blocks(inode, max_blocks);
5019 	}
5020 	while (ret >= 0 && ret < max_blocks) {
5021 		map.m_lblk += ret;
5022 		map.m_len = (max_blocks -= ret);
5023 		if (credits) {
5024 			handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
5025 						    credits);
5026 			if (IS_ERR(handle)) {
5027 				ret = PTR_ERR(handle);
5028 				break;
5029 			}
5030 		}
5031 		ret = ext4_map_blocks(handle, inode, &map,
5032 				      EXT4_GET_BLOCKS_IO_CONVERT_EXT);
5033 		if (ret <= 0)
5034 			ext4_warning(inode->i_sb,
5035 				     "inode #%lu: block %u: len %u: "
5036 				     "ext4_ext_map_blocks returned %d",
5037 				     inode->i_ino, map.m_lblk,
5038 				     map.m_len, ret);
5039 		ext4_mark_inode_dirty(handle, inode);
5040 		if (credits)
5041 			ret2 = ext4_journal_stop(handle);
5042 		if (ret <= 0 || ret2)
5043 			break;
5044 	}
5045 	if (!credits)
5046 		ret2 = ext4_journal_stop(handle);
5047 	return ret > 0 ? ret2 : ret;
5048 }
5049 
5050 /*
5051  * If newes is not existing extent (newes->ec_pblk equals zero) find
5052  * delayed extent at start of newes and update newes accordingly and
5053  * return start of the next delayed extent.
5054  *
5055  * If newes is existing extent (newes->ec_pblk is not equal zero)
5056  * return start of next delayed extent or EXT_MAX_BLOCKS if no delayed
5057  * extent found. Leave newes unmodified.
5058  */
5059 static int ext4_find_delayed_extent(struct inode *inode,
5060 				    struct extent_status *newes)
5061 {
5062 	struct extent_status es;
5063 	ext4_lblk_t block, next_del;
5064 
5065 	if (newes->es_pblk == 0) {
5066 		ext4_es_find_delayed_extent_range(inode, newes->es_lblk,
5067 				newes->es_lblk + newes->es_len - 1, &es);
5068 
5069 		/*
5070 		 * No extent in extent-tree contains block @newes->es_pblk,
5071 		 * then the block may stay in 1)a hole or 2)delayed-extent.
5072 		 */
5073 		if (es.es_len == 0)
5074 			/* A hole found. */
5075 			return 0;
5076 
5077 		if (es.es_lblk > newes->es_lblk) {
5078 			/* A hole found. */
5079 			newes->es_len = min(es.es_lblk - newes->es_lblk,
5080 					    newes->es_len);
5081 			return 0;
5082 		}
5083 
5084 		newes->es_len = es.es_lblk + es.es_len - newes->es_lblk;
5085 	}
5086 
5087 	block = newes->es_lblk + newes->es_len;
5088 	ext4_es_find_delayed_extent_range(inode, block, EXT_MAX_BLOCKS, &es);
5089 	if (es.es_len == 0)
5090 		next_del = EXT_MAX_BLOCKS;
5091 	else
5092 		next_del = es.es_lblk;
5093 
5094 	return next_del;
5095 }
5096 /* fiemap flags we can handle specified here */
5097 #define EXT4_FIEMAP_FLAGS	(FIEMAP_FLAG_SYNC|FIEMAP_FLAG_XATTR)
5098 
5099 static int ext4_xattr_fiemap(struct inode *inode,
5100 				struct fiemap_extent_info *fieinfo)
5101 {
5102 	__u64 physical = 0;
5103 	__u64 length;
5104 	__u32 flags = FIEMAP_EXTENT_LAST;
5105 	int blockbits = inode->i_sb->s_blocksize_bits;
5106 	int error = 0;
5107 
5108 	/* in-inode? */
5109 	if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
5110 		struct ext4_iloc iloc;
5111 		int offset;	/* offset of xattr in inode */
5112 
5113 		error = ext4_get_inode_loc(inode, &iloc);
5114 		if (error)
5115 			return error;
5116 		physical = (__u64)iloc.bh->b_blocknr << blockbits;
5117 		offset = EXT4_GOOD_OLD_INODE_SIZE +
5118 				EXT4_I(inode)->i_extra_isize;
5119 		physical += offset;
5120 		length = EXT4_SB(inode->i_sb)->s_inode_size - offset;
5121 		flags |= FIEMAP_EXTENT_DATA_INLINE;
5122 		brelse(iloc.bh);
5123 	} else { /* external block */
5124 		physical = (__u64)EXT4_I(inode)->i_file_acl << blockbits;
5125 		length = inode->i_sb->s_blocksize;
5126 	}
5127 
5128 	if (physical)
5129 		error = fiemap_fill_next_extent(fieinfo, 0, physical,
5130 						length, flags);
5131 	return (error < 0 ? error : 0);
5132 }
5133 
5134 int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
5135 		__u64 start, __u64 len)
5136 {
5137 	ext4_lblk_t start_blk;
5138 	int error = 0;
5139 
5140 	if (ext4_has_inline_data(inode)) {
5141 		int has_inline = 1;
5142 
5143 		error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
5144 
5145 		if (has_inline)
5146 			return error;
5147 	}
5148 
5149 	if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
5150 		error = ext4_ext_precache(inode);
5151 		if (error)
5152 			return error;
5153 	}
5154 
5155 	/* fallback to generic here if not in extents fmt */
5156 	if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
5157 		return generic_block_fiemap(inode, fieinfo, start, len,
5158 			ext4_get_block);
5159 
5160 	if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS))
5161 		return -EBADR;
5162 
5163 	if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) {
5164 		error = ext4_xattr_fiemap(inode, fieinfo);
5165 	} else {
5166 		ext4_lblk_t len_blks;
5167 		__u64 last_blk;
5168 
5169 		start_blk = start >> inode->i_sb->s_blocksize_bits;
5170 		last_blk = (start + len - 1) >> inode->i_sb->s_blocksize_bits;
5171 		if (last_blk >= EXT_MAX_BLOCKS)
5172 			last_blk = EXT_MAX_BLOCKS-1;
5173 		len_blks = ((ext4_lblk_t) last_blk) - start_blk + 1;
5174 
5175 		/*
5176 		 * Walk the extent tree gathering extent information
5177 		 * and pushing extents back to the user.
5178 		 */
5179 		error = ext4_fill_fiemap_extents(inode, start_blk,
5180 						 len_blks, fieinfo);
5181 	}
5182 	ext4_es_lru_add(inode);
5183 	return error;
5184 }
5185 
5186 /*
5187  * ext4_access_path:
5188  * Function to access the path buffer for marking it dirty.
5189  * It also checks if there are sufficient credits left in the journal handle
5190  * to update path.
5191  */
5192 static int
5193 ext4_access_path(handle_t *handle, struct inode *inode,
5194 		struct ext4_ext_path *path)
5195 {
5196 	int credits, err;
5197 
5198 	if (!ext4_handle_valid(handle))
5199 		return 0;
5200 
5201 	/*
5202 	 * Check if need to extend journal credits
5203 	 * 3 for leaf, sb, and inode plus 2 (bmap and group
5204 	 * descriptor) for each block group; assume two block
5205 	 * groups
5206 	 */
5207 	if (handle->h_buffer_credits < 7) {
5208 		credits = ext4_writepage_trans_blocks(inode);
5209 		err = ext4_ext_truncate_extend_restart(handle, inode, credits);
5210 		/* EAGAIN is success */
5211 		if (err && err != -EAGAIN)
5212 			return err;
5213 	}
5214 
5215 	err = ext4_ext_get_access(handle, inode, path);
5216 	return err;
5217 }
5218 
5219 /*
5220  * ext4_ext_shift_path_extents:
5221  * Shift the extents of a path structure lying between path[depth].p_ext
5222  * and EXT_LAST_EXTENT(path[depth].p_hdr) downwards, by subtracting shift
5223  * from starting block for each extent.
5224  */
5225 static int
5226 ext4_ext_shift_path_extents(struct ext4_ext_path *path, ext4_lblk_t shift,
5227 			    struct inode *inode, handle_t *handle,
5228 			    ext4_lblk_t *start)
5229 {
5230 	int depth, err = 0;
5231 	struct ext4_extent *ex_start, *ex_last;
5232 	bool update = 0;
5233 	depth = path->p_depth;
5234 
5235 	while (depth >= 0) {
5236 		if (depth == path->p_depth) {
5237 			ex_start = path[depth].p_ext;
5238 			if (!ex_start)
5239 				return -EIO;
5240 
5241 			ex_last = EXT_LAST_EXTENT(path[depth].p_hdr);
5242 			if (!ex_last)
5243 				return -EIO;
5244 
5245 			err = ext4_access_path(handle, inode, path + depth);
5246 			if (err)
5247 				goto out;
5248 
5249 			if (ex_start == EXT_FIRST_EXTENT(path[depth].p_hdr))
5250 				update = 1;
5251 
5252 			*start = le32_to_cpu(ex_last->ee_block) +
5253 				ext4_ext_get_actual_len(ex_last);
5254 
5255 			while (ex_start <= ex_last) {
5256 				le32_add_cpu(&ex_start->ee_block, -shift);
5257 				/* Try to merge to the left. */
5258 				if ((ex_start >
5259 				     EXT_FIRST_EXTENT(path[depth].p_hdr)) &&
5260 				    ext4_ext_try_to_merge_right(inode,
5261 							path, ex_start - 1))
5262 					ex_last--;
5263 				else
5264 					ex_start++;
5265 			}
5266 			err = ext4_ext_dirty(handle, inode, path + depth);
5267 			if (err)
5268 				goto out;
5269 
5270 			if (--depth < 0 || !update)
5271 				break;
5272 		}
5273 
5274 		/* Update index too */
5275 		err = ext4_access_path(handle, inode, path + depth);
5276 		if (err)
5277 			goto out;
5278 
5279 		le32_add_cpu(&path[depth].p_idx->ei_block, -shift);
5280 		err = ext4_ext_dirty(handle, inode, path + depth);
5281 		if (err)
5282 			goto out;
5283 
5284 		/* we are done if current index is not a starting index */
5285 		if (path[depth].p_idx != EXT_FIRST_INDEX(path[depth].p_hdr))
5286 			break;
5287 
5288 		depth--;
5289 	}
5290 
5291 out:
5292 	return err;
5293 }
5294 
5295 /*
5296  * ext4_ext_shift_extents:
5297  * All the extents which lies in the range from start to the last allocated
5298  * block for the file are shifted downwards by shift blocks.
5299  * On success, 0 is returned, error otherwise.
5300  */
5301 static int
5302 ext4_ext_shift_extents(struct inode *inode, handle_t *handle,
5303 		       ext4_lblk_t start, ext4_lblk_t shift)
5304 {
5305 	struct ext4_ext_path *path;
5306 	int ret = 0, depth;
5307 	struct ext4_extent *extent;
5308 	ext4_lblk_t stop_block;
5309 	ext4_lblk_t ex_start, ex_end;
5310 
5311 	/* Let path point to the last extent */
5312 	path = ext4_find_extent(inode, EXT_MAX_BLOCKS - 1, NULL, 0);
5313 	if (IS_ERR(path))
5314 		return PTR_ERR(path);
5315 
5316 	depth = path->p_depth;
5317 	extent = path[depth].p_ext;
5318 	if (!extent)
5319 		goto out;
5320 
5321 	stop_block = le32_to_cpu(extent->ee_block) +
5322 			ext4_ext_get_actual_len(extent);
5323 
5324 	/* Nothing to shift, if hole is at the end of file */
5325 	if (start >= stop_block)
5326 		goto out;
5327 
5328 	/*
5329 	 * Don't start shifting extents until we make sure the hole is big
5330 	 * enough to accomodate the shift.
5331 	 */
5332 	path = ext4_find_extent(inode, start - 1, &path, 0);
5333 	if (IS_ERR(path))
5334 		return PTR_ERR(path);
5335 	depth = path->p_depth;
5336 	extent =  path[depth].p_ext;
5337 	if (extent) {
5338 		ex_start = le32_to_cpu(extent->ee_block);
5339 		ex_end = le32_to_cpu(extent->ee_block) +
5340 			ext4_ext_get_actual_len(extent);
5341 	} else {
5342 		ex_start = 0;
5343 		ex_end = 0;
5344 	}
5345 
5346 	if ((start == ex_start && shift > ex_start) ||
5347 	    (shift > start - ex_end))
5348 		return -EINVAL;
5349 
5350 	/* Its safe to start updating extents */
5351 	while (start < stop_block) {
5352 		path = ext4_find_extent(inode, start, &path, 0);
5353 		if (IS_ERR(path))
5354 			return PTR_ERR(path);
5355 		depth = path->p_depth;
5356 		extent = path[depth].p_ext;
5357 		if (!extent) {
5358 			EXT4_ERROR_INODE(inode, "unexpected hole at %lu",
5359 					 (unsigned long) start);
5360 			return -EIO;
5361 		}
5362 		if (start > le32_to_cpu(extent->ee_block)) {
5363 			/* Hole, move to the next extent */
5364 			if (extent < EXT_LAST_EXTENT(path[depth].p_hdr)) {
5365 				path[depth].p_ext++;
5366 			} else {
5367 				start = ext4_ext_next_allocated_block(path);
5368 				continue;
5369 			}
5370 		}
5371 		ret = ext4_ext_shift_path_extents(path, shift, inode,
5372 				handle, &start);
5373 		if (ret)
5374 			break;
5375 	}
5376 out:
5377 	ext4_ext_drop_refs(path);
5378 	kfree(path);
5379 	return ret;
5380 }
5381 
5382 /*
5383  * ext4_collapse_range:
5384  * This implements the fallocate's collapse range functionality for ext4
5385  * Returns: 0 and non-zero on error.
5386  */
5387 int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
5388 {
5389 	struct super_block *sb = inode->i_sb;
5390 	ext4_lblk_t punch_start, punch_stop;
5391 	handle_t *handle;
5392 	unsigned int credits;
5393 	loff_t new_size, ioffset;
5394 	int ret;
5395 
5396 	/* Collapse range works only on fs block size aligned offsets. */
5397 	if (offset & (EXT4_CLUSTER_SIZE(sb) - 1) ||
5398 	    len & (EXT4_CLUSTER_SIZE(sb) - 1))
5399 		return -EINVAL;
5400 
5401 	if (!S_ISREG(inode->i_mode))
5402 		return -EINVAL;
5403 
5404 	trace_ext4_collapse_range(inode, offset, len);
5405 
5406 	punch_start = offset >> EXT4_BLOCK_SIZE_BITS(sb);
5407 	punch_stop = (offset + len) >> EXT4_BLOCK_SIZE_BITS(sb);
5408 
5409 	/* Call ext4_force_commit to flush all data in case of data=journal. */
5410 	if (ext4_should_journal_data(inode)) {
5411 		ret = ext4_force_commit(inode->i_sb);
5412 		if (ret)
5413 			return ret;
5414 	}
5415 
5416 	/*
5417 	 * Need to round down offset to be aligned with page size boundary
5418 	 * for page size > block size.
5419 	 */
5420 	ioffset = round_down(offset, PAGE_SIZE);
5421 
5422 	/* Write out all dirty pages */
5423 	ret = filemap_write_and_wait_range(inode->i_mapping, ioffset,
5424 					   LLONG_MAX);
5425 	if (ret)
5426 		return ret;
5427 
5428 	/* Take mutex lock */
5429 	mutex_lock(&inode->i_mutex);
5430 
5431 	/*
5432 	 * There is no need to overlap collapse range with EOF, in which case
5433 	 * it is effectively a truncate operation
5434 	 */
5435 	if (offset + len >= i_size_read(inode)) {
5436 		ret = -EINVAL;
5437 		goto out_mutex;
5438 	}
5439 
5440 	/* Currently just for extent based files */
5441 	if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
5442 		ret = -EOPNOTSUPP;
5443 		goto out_mutex;
5444 	}
5445 
5446 	truncate_pagecache(inode, ioffset);
5447 
5448 	/* Wait for existing dio to complete */
5449 	ext4_inode_block_unlocked_dio(inode);
5450 	inode_dio_wait(inode);
5451 
5452 	credits = ext4_writepage_trans_blocks(inode);
5453 	handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
5454 	if (IS_ERR(handle)) {
5455 		ret = PTR_ERR(handle);
5456 		goto out_dio;
5457 	}
5458 
5459 	down_write(&EXT4_I(inode)->i_data_sem);
5460 	ext4_discard_preallocations(inode);
5461 
5462 	ret = ext4_es_remove_extent(inode, punch_start,
5463 				    EXT_MAX_BLOCKS - punch_start);
5464 	if (ret) {
5465 		up_write(&EXT4_I(inode)->i_data_sem);
5466 		goto out_stop;
5467 	}
5468 
5469 	ret = ext4_ext_remove_space(inode, punch_start, punch_stop - 1);
5470 	if (ret) {
5471 		up_write(&EXT4_I(inode)->i_data_sem);
5472 		goto out_stop;
5473 	}
5474 	ext4_discard_preallocations(inode);
5475 
5476 	ret = ext4_ext_shift_extents(inode, handle, punch_stop,
5477 				     punch_stop - punch_start);
5478 	if (ret) {
5479 		up_write(&EXT4_I(inode)->i_data_sem);
5480 		goto out_stop;
5481 	}
5482 
5483 	new_size = i_size_read(inode) - len;
5484 	i_size_write(inode, new_size);
5485 	EXT4_I(inode)->i_disksize = new_size;
5486 
5487 	up_write(&EXT4_I(inode)->i_data_sem);
5488 	if (IS_SYNC(inode))
5489 		ext4_handle_sync(handle);
5490 	inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
5491 	ext4_mark_inode_dirty(handle, inode);
5492 
5493 out_stop:
5494 	ext4_journal_stop(handle);
5495 out_dio:
5496 	ext4_inode_resume_unlocked_dio(inode);
5497 out_mutex:
5498 	mutex_unlock(&inode->i_mutex);
5499 	return ret;
5500 }
5501 
5502 /**
5503  * ext4_swap_extents - Swap extents between two inodes
5504  *
5505  * @inode1:	First inode
5506  * @inode2:	Second inode
5507  * @lblk1:	Start block for first inode
5508  * @lblk2:	Start block for second inode
5509  * @count:	Number of blocks to swap
5510  * @mark_unwritten: Mark second inode's extents as unwritten after swap
5511  * @erp:	Pointer to save error value
5512  *
5513  * This helper routine does exactly what is promise "swap extents". All other
5514  * stuff such as page-cache locking consistency, bh mapping consistency or
5515  * extent's data copying must be performed by caller.
5516  * Locking:
5517  * 		i_mutex is held for both inodes
5518  * 		i_data_sem is locked for write for both inodes
5519  * Assumptions:
5520  *		All pages from requested range are locked for both inodes
5521  */
5522 int
5523 ext4_swap_extents(handle_t *handle, struct inode *inode1,
5524 		     struct inode *inode2, ext4_lblk_t lblk1, ext4_lblk_t lblk2,
5525 		  ext4_lblk_t count, int unwritten, int *erp)
5526 {
5527 	struct ext4_ext_path *path1 = NULL;
5528 	struct ext4_ext_path *path2 = NULL;
5529 	int replaced_count = 0;
5530 
5531 	BUG_ON(!rwsem_is_locked(&EXT4_I(inode1)->i_data_sem));
5532 	BUG_ON(!rwsem_is_locked(&EXT4_I(inode2)->i_data_sem));
5533 	BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5534 	BUG_ON(!mutex_is_locked(&inode1->i_mutex));
5535 
5536 	*erp = ext4_es_remove_extent(inode1, lblk1, count);
5537 	if (unlikely(*erp))
5538 		return 0;
5539 	*erp = ext4_es_remove_extent(inode2, lblk2, count);
5540 	if (unlikely(*erp))
5541 		return 0;
5542 
5543 	while (count) {
5544 		struct ext4_extent *ex1, *ex2, tmp_ex;
5545 		ext4_lblk_t e1_blk, e2_blk;
5546 		int e1_len, e2_len, len;
5547 		int split = 0;
5548 
5549 		path1 = ext4_find_extent(inode1, lblk1, NULL, EXT4_EX_NOCACHE);
5550 		if (unlikely(IS_ERR(path1))) {
5551 			*erp = PTR_ERR(path1);
5552 			path1 = NULL;
5553 		finish:
5554 			count = 0;
5555 			goto repeat;
5556 		}
5557 		path2 = ext4_find_extent(inode2, lblk2, NULL, EXT4_EX_NOCACHE);
5558 		if (unlikely(IS_ERR(path2))) {
5559 			*erp = PTR_ERR(path2);
5560 			path2 = NULL;
5561 			goto finish;
5562 		}
5563 		ex1 = path1[path1->p_depth].p_ext;
5564 		ex2 = path2[path2->p_depth].p_ext;
5565 		/* Do we have somthing to swap ? */
5566 		if (unlikely(!ex2 || !ex1))
5567 			goto finish;
5568 
5569 		e1_blk = le32_to_cpu(ex1->ee_block);
5570 		e2_blk = le32_to_cpu(ex2->ee_block);
5571 		e1_len = ext4_ext_get_actual_len(ex1);
5572 		e2_len = ext4_ext_get_actual_len(ex2);
5573 
5574 		/* Hole handling */
5575 		if (!in_range(lblk1, e1_blk, e1_len) ||
5576 		    !in_range(lblk2, e2_blk, e2_len)) {
5577 			ext4_lblk_t next1, next2;
5578 
5579 			/* if hole after extent, then go to next extent */
5580 			next1 = ext4_ext_next_allocated_block(path1);
5581 			next2 = ext4_ext_next_allocated_block(path2);
5582 			/* If hole before extent, then shift to that extent */
5583 			if (e1_blk > lblk1)
5584 				next1 = e1_blk;
5585 			if (e2_blk > lblk2)
5586 				next2 = e1_blk;
5587 			/* Do we have something to swap */
5588 			if (next1 == EXT_MAX_BLOCKS || next2 == EXT_MAX_BLOCKS)
5589 				goto finish;
5590 			/* Move to the rightest boundary */
5591 			len = next1 - lblk1;
5592 			if (len < next2 - lblk2)
5593 				len = next2 - lblk2;
5594 			if (len > count)
5595 				len = count;
5596 			lblk1 += len;
5597 			lblk2 += len;
5598 			count -= len;
5599 			goto repeat;
5600 		}
5601 
5602 		/* Prepare left boundary */
5603 		if (e1_blk < lblk1) {
5604 			split = 1;
5605 			*erp = ext4_force_split_extent_at(handle, inode1,
5606 						&path1, lblk1, 0);
5607 			if (unlikely(*erp))
5608 				goto finish;
5609 		}
5610 		if (e2_blk < lblk2) {
5611 			split = 1;
5612 			*erp = ext4_force_split_extent_at(handle, inode2,
5613 						&path2,  lblk2, 0);
5614 			if (unlikely(*erp))
5615 				goto finish;
5616 		}
5617 		/* ext4_split_extent_at() may result in leaf extent split,
5618 		 * path must to be revalidated. */
5619 		if (split)
5620 			goto repeat;
5621 
5622 		/* Prepare right boundary */
5623 		len = count;
5624 		if (len > e1_blk + e1_len - lblk1)
5625 			len = e1_blk + e1_len - lblk1;
5626 		if (len > e2_blk + e2_len - lblk2)
5627 			len = e2_blk + e2_len - lblk2;
5628 
5629 		if (len != e1_len) {
5630 			split = 1;
5631 			*erp = ext4_force_split_extent_at(handle, inode1,
5632 						&path1, lblk1 + len, 0);
5633 			if (unlikely(*erp))
5634 				goto finish;
5635 		}
5636 		if (len != e2_len) {
5637 			split = 1;
5638 			*erp = ext4_force_split_extent_at(handle, inode2,
5639 						&path2, lblk2 + len, 0);
5640 			if (*erp)
5641 				goto finish;
5642 		}
5643 		/* ext4_split_extent_at() may result in leaf extent split,
5644 		 * path must to be revalidated. */
5645 		if (split)
5646 			goto repeat;
5647 
5648 		BUG_ON(e2_len != e1_len);
5649 		*erp = ext4_ext_get_access(handle, inode1, path1 + path1->p_depth);
5650 		if (unlikely(*erp))
5651 			goto finish;
5652 		*erp = ext4_ext_get_access(handle, inode2, path2 + path2->p_depth);
5653 		if (unlikely(*erp))
5654 			goto finish;
5655 
5656 		/* Both extents are fully inside boundaries. Swap it now */
5657 		tmp_ex = *ex1;
5658 		ext4_ext_store_pblock(ex1, ext4_ext_pblock(ex2));
5659 		ext4_ext_store_pblock(ex2, ext4_ext_pblock(&tmp_ex));
5660 		ex1->ee_len = cpu_to_le16(e2_len);
5661 		ex2->ee_len = cpu_to_le16(e1_len);
5662 		if (unwritten)
5663 			ext4_ext_mark_unwritten(ex2);
5664 		if (ext4_ext_is_unwritten(&tmp_ex))
5665 			ext4_ext_mark_unwritten(ex1);
5666 
5667 		ext4_ext_try_to_merge(handle, inode2, path2, ex2);
5668 		ext4_ext_try_to_merge(handle, inode1, path1, ex1);
5669 		*erp = ext4_ext_dirty(handle, inode2, path2 +
5670 				      path2->p_depth);
5671 		if (unlikely(*erp))
5672 			goto finish;
5673 		*erp = ext4_ext_dirty(handle, inode1, path1 +
5674 				      path1->p_depth);
5675 		/*
5676 		 * Looks scarry ah..? second inode already points to new blocks,
5677 		 * and it was successfully dirtied. But luckily error may happen
5678 		 * only due to journal error, so full transaction will be
5679 		 * aborted anyway.
5680 		 */
5681 		if (unlikely(*erp))
5682 			goto finish;
5683 		lblk1 += len;
5684 		lblk2 += len;
5685 		replaced_count += len;
5686 		count -= len;
5687 
5688 	repeat:
5689 		ext4_ext_drop_refs(path1);
5690 		kfree(path1);
5691 		ext4_ext_drop_refs(path2);
5692 		kfree(path2);
5693 		path1 = path2 = NULL;
5694 	}
5695 	return replaced_count;
5696 }
5697