xref: /linux/fs/xfs/libxfs/xfs_iext_tree.c (revision 44f57d78)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2017 Christoph Hellwig.
4  */
5 
6 #include <linux/cache.h>
7 #include <linux/kernel.h>
8 #include <linux/slab.h>
9 #include "xfs.h"
10 #include "xfs_format.h"
11 #include "xfs_bit.h"
12 #include "xfs_log_format.h"
13 #include "xfs_inode.h"
14 #include "xfs_inode_fork.h"
15 #include "xfs_trans_resv.h"
16 #include "xfs_mount.h"
17 #include "xfs_bmap.h"
18 #include "xfs_trace.h"
19 
20 /*
21  * In-core extent record layout:
22  *
23  * +-------+----------------------------+
24  * | 00:53 | all 54 bits of startoff    |
25  * | 54:63 | low 10 bits of startblock  |
26  * +-------+----------------------------+
27  * | 00:20 | all 21 bits of length      |
28  * |    21 | unwritten extent bit       |
29  * | 22:63 | high 42 bits of startblock |
30  * +-------+----------------------------+
31  */
32 #define XFS_IEXT_STARTOFF_MASK		xfs_mask64lo(BMBT_STARTOFF_BITLEN)
33 #define XFS_IEXT_LENGTH_MASK		xfs_mask64lo(BMBT_BLOCKCOUNT_BITLEN)
34 #define XFS_IEXT_STARTBLOCK_MASK	xfs_mask64lo(BMBT_STARTBLOCK_BITLEN)
35 
36 struct xfs_iext_rec {
37 	uint64_t			lo;
38 	uint64_t			hi;
39 };
40 
41 /*
42  * Given that the length can't be a zero, only an empty hi value indicates an
43  * unused record.
44  */
45 static bool xfs_iext_rec_is_empty(struct xfs_iext_rec *rec)
46 {
47 	return rec->hi == 0;
48 }
49 
50 static inline void xfs_iext_rec_clear(struct xfs_iext_rec *rec)
51 {
52 	rec->lo = 0;
53 	rec->hi = 0;
54 }
55 
56 static void
57 xfs_iext_set(
58 	struct xfs_iext_rec	*rec,
59 	struct xfs_bmbt_irec	*irec)
60 {
61 	ASSERT((irec->br_startoff & ~XFS_IEXT_STARTOFF_MASK) == 0);
62 	ASSERT((irec->br_blockcount & ~XFS_IEXT_LENGTH_MASK) == 0);
63 	ASSERT((irec->br_startblock & ~XFS_IEXT_STARTBLOCK_MASK) == 0);
64 
65 	rec->lo = irec->br_startoff & XFS_IEXT_STARTOFF_MASK;
66 	rec->hi = irec->br_blockcount & XFS_IEXT_LENGTH_MASK;
67 
68 	rec->lo |= (irec->br_startblock << 54);
69 	rec->hi |= ((irec->br_startblock & ~xfs_mask64lo(10)) << (22 - 10));
70 
71 	if (irec->br_state == XFS_EXT_UNWRITTEN)
72 		rec->hi |= (1 << 21);
73 }
74 
75 static void
76 xfs_iext_get(
77 	struct xfs_bmbt_irec	*irec,
78 	struct xfs_iext_rec	*rec)
79 {
80 	irec->br_startoff = rec->lo & XFS_IEXT_STARTOFF_MASK;
81 	irec->br_blockcount = rec->hi & XFS_IEXT_LENGTH_MASK;
82 
83 	irec->br_startblock = rec->lo >> 54;
84 	irec->br_startblock |= (rec->hi & xfs_mask64hi(42)) >> (22 - 10);
85 
86 	if (rec->hi & (1 << 21))
87 		irec->br_state = XFS_EXT_UNWRITTEN;
88 	else
89 		irec->br_state = XFS_EXT_NORM;
90 }
91 
92 enum {
93 	NODE_SIZE	= 256,
94 	KEYS_PER_NODE	= NODE_SIZE / (sizeof(uint64_t) + sizeof(void *)),
95 	RECS_PER_LEAF	= (NODE_SIZE - (2 * sizeof(struct xfs_iext_leaf *))) /
96 				sizeof(struct xfs_iext_rec),
97 };
98 
99 /*
100  * In-core extent btree block layout:
101  *
102  * There are two types of blocks in the btree: leaf and inner (non-leaf) blocks.
103  *
104  * The leaf blocks are made up by %KEYS_PER_NODE extent records, which each
105  * contain the startoffset, blockcount, startblock and unwritten extent flag.
106  * See above for the exact format, followed by pointers to the previous and next
107  * leaf blocks (if there are any).
108  *
109  * The inner (non-leaf) blocks first contain KEYS_PER_NODE lookup keys, followed
110  * by an equal number of pointers to the btree blocks at the next lower level.
111  *
112  *		+-------+-------+-------+-------+-------+----------+----------+
113  * Leaf:	| rec 1 | rec 2 | rec 3 | rec 4 | rec N | prev-ptr | next-ptr |
114  *		+-------+-------+-------+-------+-------+----------+----------+
115  *
116  *		+-------+-------+-------+-------+-------+-------+------+-------+
117  * Inner:	| key 1 | key 2 | key 3 | key N | ptr 1 | ptr 2 | ptr3 | ptr N |
118  *		+-------+-------+-------+-------+-------+-------+------+-------+
119  */
120 struct xfs_iext_node {
121 	uint64_t		keys[KEYS_PER_NODE];
122 #define XFS_IEXT_KEY_INVALID	(1ULL << 63)
123 	void			*ptrs[KEYS_PER_NODE];
124 };
125 
126 struct xfs_iext_leaf {
127 	struct xfs_iext_rec	recs[RECS_PER_LEAF];
128 	struct xfs_iext_leaf	*prev;
129 	struct xfs_iext_leaf	*next;
130 };
131 
132 inline xfs_extnum_t xfs_iext_count(struct xfs_ifork *ifp)
133 {
134 	return ifp->if_bytes / sizeof(struct xfs_iext_rec);
135 }
136 
137 static inline int xfs_iext_max_recs(struct xfs_ifork *ifp)
138 {
139 	if (ifp->if_height == 1)
140 		return xfs_iext_count(ifp);
141 	return RECS_PER_LEAF;
142 }
143 
144 static inline struct xfs_iext_rec *cur_rec(struct xfs_iext_cursor *cur)
145 {
146 	return &cur->leaf->recs[cur->pos];
147 }
148 
149 static inline bool xfs_iext_valid(struct xfs_ifork *ifp,
150 		struct xfs_iext_cursor *cur)
151 {
152 	if (!cur->leaf)
153 		return false;
154 	if (cur->pos < 0 || cur->pos >= xfs_iext_max_recs(ifp))
155 		return false;
156 	if (xfs_iext_rec_is_empty(cur_rec(cur)))
157 		return false;
158 	return true;
159 }
160 
161 static void *
162 xfs_iext_find_first_leaf(
163 	struct xfs_ifork	*ifp)
164 {
165 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
166 	int			height;
167 
168 	if (!ifp->if_height)
169 		return NULL;
170 
171 	for (height = ifp->if_height; height > 1; height--) {
172 		node = node->ptrs[0];
173 		ASSERT(node);
174 	}
175 
176 	return node;
177 }
178 
179 static void *
180 xfs_iext_find_last_leaf(
181 	struct xfs_ifork	*ifp)
182 {
183 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
184 	int			height, i;
185 
186 	if (!ifp->if_height)
187 		return NULL;
188 
189 	for (height = ifp->if_height; height > 1; height--) {
190 		for (i = 1; i < KEYS_PER_NODE; i++)
191 			if (!node->ptrs[i])
192 				break;
193 		node = node->ptrs[i - 1];
194 		ASSERT(node);
195 	}
196 
197 	return node;
198 }
199 
200 void
201 xfs_iext_first(
202 	struct xfs_ifork	*ifp,
203 	struct xfs_iext_cursor	*cur)
204 {
205 	cur->pos = 0;
206 	cur->leaf = xfs_iext_find_first_leaf(ifp);
207 }
208 
209 void
210 xfs_iext_last(
211 	struct xfs_ifork	*ifp,
212 	struct xfs_iext_cursor	*cur)
213 {
214 	int			i;
215 
216 	cur->leaf = xfs_iext_find_last_leaf(ifp);
217 	if (!cur->leaf) {
218 		cur->pos = 0;
219 		return;
220 	}
221 
222 	for (i = 1; i < xfs_iext_max_recs(ifp); i++) {
223 		if (xfs_iext_rec_is_empty(&cur->leaf->recs[i]))
224 			break;
225 	}
226 	cur->pos = i - 1;
227 }
228 
229 void
230 xfs_iext_next(
231 	struct xfs_ifork	*ifp,
232 	struct xfs_iext_cursor	*cur)
233 {
234 	if (!cur->leaf) {
235 		ASSERT(cur->pos <= 0 || cur->pos >= RECS_PER_LEAF);
236 		xfs_iext_first(ifp, cur);
237 		return;
238 	}
239 
240 	ASSERT(cur->pos >= 0);
241 	ASSERT(cur->pos < xfs_iext_max_recs(ifp));
242 
243 	cur->pos++;
244 	if (ifp->if_height > 1 && !xfs_iext_valid(ifp, cur) &&
245 	    cur->leaf->next) {
246 		cur->leaf = cur->leaf->next;
247 		cur->pos = 0;
248 	}
249 }
250 
251 void
252 xfs_iext_prev(
253 	struct xfs_ifork	*ifp,
254 	struct xfs_iext_cursor	*cur)
255 {
256 	if (!cur->leaf) {
257 		ASSERT(cur->pos <= 0 || cur->pos >= RECS_PER_LEAF);
258 		xfs_iext_last(ifp, cur);
259 		return;
260 	}
261 
262 	ASSERT(cur->pos >= 0);
263 	ASSERT(cur->pos <= RECS_PER_LEAF);
264 
265 recurse:
266 	do {
267 		cur->pos--;
268 		if (xfs_iext_valid(ifp, cur))
269 			return;
270 	} while (cur->pos > 0);
271 
272 	if (ifp->if_height > 1 && cur->leaf->prev) {
273 		cur->leaf = cur->leaf->prev;
274 		cur->pos = RECS_PER_LEAF;
275 		goto recurse;
276 	}
277 }
278 
279 static inline int
280 xfs_iext_key_cmp(
281 	struct xfs_iext_node	*node,
282 	int			n,
283 	xfs_fileoff_t		offset)
284 {
285 	if (node->keys[n] > offset)
286 		return 1;
287 	if (node->keys[n] < offset)
288 		return -1;
289 	return 0;
290 }
291 
292 static inline int
293 xfs_iext_rec_cmp(
294 	struct xfs_iext_rec	*rec,
295 	xfs_fileoff_t		offset)
296 {
297 	uint64_t		rec_offset = rec->lo & XFS_IEXT_STARTOFF_MASK;
298 	uint32_t		rec_len = rec->hi & XFS_IEXT_LENGTH_MASK;
299 
300 	if (rec_offset > offset)
301 		return 1;
302 	if (rec_offset + rec_len <= offset)
303 		return -1;
304 	return 0;
305 }
306 
307 static void *
308 xfs_iext_find_level(
309 	struct xfs_ifork	*ifp,
310 	xfs_fileoff_t		offset,
311 	int			level)
312 {
313 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
314 	int			height, i;
315 
316 	if (!ifp->if_height)
317 		return NULL;
318 
319 	for (height = ifp->if_height; height > level; height--) {
320 		for (i = 1; i < KEYS_PER_NODE; i++)
321 			if (xfs_iext_key_cmp(node, i, offset) > 0)
322 				break;
323 
324 		node = node->ptrs[i - 1];
325 		if (!node)
326 			break;
327 	}
328 
329 	return node;
330 }
331 
332 static int
333 xfs_iext_node_pos(
334 	struct xfs_iext_node	*node,
335 	xfs_fileoff_t		offset)
336 {
337 	int			i;
338 
339 	for (i = 1; i < KEYS_PER_NODE; i++) {
340 		if (xfs_iext_key_cmp(node, i, offset) > 0)
341 			break;
342 	}
343 
344 	return i - 1;
345 }
346 
347 static int
348 xfs_iext_node_insert_pos(
349 	struct xfs_iext_node	*node,
350 	xfs_fileoff_t		offset)
351 {
352 	int			i;
353 
354 	for (i = 0; i < KEYS_PER_NODE; i++) {
355 		if (xfs_iext_key_cmp(node, i, offset) > 0)
356 			return i;
357 	}
358 
359 	return KEYS_PER_NODE;
360 }
361 
362 static int
363 xfs_iext_node_nr_entries(
364 	struct xfs_iext_node	*node,
365 	int			start)
366 {
367 	int			i;
368 
369 	for (i = start; i < KEYS_PER_NODE; i++) {
370 		if (node->keys[i] == XFS_IEXT_KEY_INVALID)
371 			break;
372 	}
373 
374 	return i;
375 }
376 
377 static int
378 xfs_iext_leaf_nr_entries(
379 	struct xfs_ifork	*ifp,
380 	struct xfs_iext_leaf	*leaf,
381 	int			start)
382 {
383 	int			i;
384 
385 	for (i = start; i < xfs_iext_max_recs(ifp); i++) {
386 		if (xfs_iext_rec_is_empty(&leaf->recs[i]))
387 			break;
388 	}
389 
390 	return i;
391 }
392 
393 static inline uint64_t
394 xfs_iext_leaf_key(
395 	struct xfs_iext_leaf	*leaf,
396 	int			n)
397 {
398 	return leaf->recs[n].lo & XFS_IEXT_STARTOFF_MASK;
399 }
400 
401 static void
402 xfs_iext_grow(
403 	struct xfs_ifork	*ifp)
404 {
405 	struct xfs_iext_node	*node = kmem_zalloc(NODE_SIZE, KM_NOFS);
406 	int			i;
407 
408 	if (ifp->if_height == 1) {
409 		struct xfs_iext_leaf *prev = ifp->if_u1.if_root;
410 
411 		node->keys[0] = xfs_iext_leaf_key(prev, 0);
412 		node->ptrs[0] = prev;
413 	} else  {
414 		struct xfs_iext_node *prev = ifp->if_u1.if_root;
415 
416 		ASSERT(ifp->if_height > 1);
417 
418 		node->keys[0] = prev->keys[0];
419 		node->ptrs[0] = prev;
420 	}
421 
422 	for (i = 1; i < KEYS_PER_NODE; i++)
423 		node->keys[i] = XFS_IEXT_KEY_INVALID;
424 
425 	ifp->if_u1.if_root = node;
426 	ifp->if_height++;
427 }
428 
429 static void
430 xfs_iext_update_node(
431 	struct xfs_ifork	*ifp,
432 	xfs_fileoff_t		old_offset,
433 	xfs_fileoff_t		new_offset,
434 	int			level,
435 	void			*ptr)
436 {
437 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
438 	int			height, i;
439 
440 	for (height = ifp->if_height; height > level; height--) {
441 		for (i = 0; i < KEYS_PER_NODE; i++) {
442 			if (i > 0 && xfs_iext_key_cmp(node, i, old_offset) > 0)
443 				break;
444 			if (node->keys[i] == old_offset)
445 				node->keys[i] = new_offset;
446 		}
447 		node = node->ptrs[i - 1];
448 		ASSERT(node);
449 	}
450 
451 	ASSERT(node == ptr);
452 }
453 
454 static struct xfs_iext_node *
455 xfs_iext_split_node(
456 	struct xfs_iext_node	**nodep,
457 	int			*pos,
458 	int			*nr_entries)
459 {
460 	struct xfs_iext_node	*node = *nodep;
461 	struct xfs_iext_node	*new = kmem_zalloc(NODE_SIZE, KM_NOFS);
462 	const int		nr_move = KEYS_PER_NODE / 2;
463 	int			nr_keep = nr_move + (KEYS_PER_NODE & 1);
464 	int			i = 0;
465 
466 	/* for sequential append operations just spill over into the new node */
467 	if (*pos == KEYS_PER_NODE) {
468 		*nodep = new;
469 		*pos = 0;
470 		*nr_entries = 0;
471 		goto done;
472 	}
473 
474 
475 	for (i = 0; i < nr_move; i++) {
476 		new->keys[i] = node->keys[nr_keep + i];
477 		new->ptrs[i] = node->ptrs[nr_keep + i];
478 
479 		node->keys[nr_keep + i] = XFS_IEXT_KEY_INVALID;
480 		node->ptrs[nr_keep + i] = NULL;
481 	}
482 
483 	if (*pos >= nr_keep) {
484 		*nodep = new;
485 		*pos -= nr_keep;
486 		*nr_entries = nr_move;
487 	} else {
488 		*nr_entries = nr_keep;
489 	}
490 done:
491 	for (; i < KEYS_PER_NODE; i++)
492 		new->keys[i] = XFS_IEXT_KEY_INVALID;
493 	return new;
494 }
495 
496 static void
497 xfs_iext_insert_node(
498 	struct xfs_ifork	*ifp,
499 	uint64_t		offset,
500 	void			*ptr,
501 	int			level)
502 {
503 	struct xfs_iext_node	*node, *new;
504 	int			i, pos, nr_entries;
505 
506 again:
507 	if (ifp->if_height < level)
508 		xfs_iext_grow(ifp);
509 
510 	new = NULL;
511 	node = xfs_iext_find_level(ifp, offset, level);
512 	pos = xfs_iext_node_insert_pos(node, offset);
513 	nr_entries = xfs_iext_node_nr_entries(node, pos);
514 
515 	ASSERT(pos >= nr_entries || xfs_iext_key_cmp(node, pos, offset) != 0);
516 	ASSERT(nr_entries <= KEYS_PER_NODE);
517 
518 	if (nr_entries == KEYS_PER_NODE)
519 		new = xfs_iext_split_node(&node, &pos, &nr_entries);
520 
521 	/*
522 	 * Update the pointers in higher levels if the first entry changes
523 	 * in an existing node.
524 	 */
525 	if (node != new && pos == 0 && nr_entries > 0)
526 		xfs_iext_update_node(ifp, node->keys[0], offset, level, node);
527 
528 	for (i = nr_entries; i > pos; i--) {
529 		node->keys[i] = node->keys[i - 1];
530 		node->ptrs[i] = node->ptrs[i - 1];
531 	}
532 	node->keys[pos] = offset;
533 	node->ptrs[pos] = ptr;
534 
535 	if (new) {
536 		offset = new->keys[0];
537 		ptr = new;
538 		level++;
539 		goto again;
540 	}
541 }
542 
543 static struct xfs_iext_leaf *
544 xfs_iext_split_leaf(
545 	struct xfs_iext_cursor	*cur,
546 	int			*nr_entries)
547 {
548 	struct xfs_iext_leaf	*leaf = cur->leaf;
549 	struct xfs_iext_leaf	*new = kmem_zalloc(NODE_SIZE, KM_NOFS);
550 	const int		nr_move = RECS_PER_LEAF / 2;
551 	int			nr_keep = nr_move + (RECS_PER_LEAF & 1);
552 	int			i;
553 
554 	/* for sequential append operations just spill over into the new node */
555 	if (cur->pos == RECS_PER_LEAF) {
556 		cur->leaf = new;
557 		cur->pos = 0;
558 		*nr_entries = 0;
559 		goto done;
560 	}
561 
562 	for (i = 0; i < nr_move; i++) {
563 		new->recs[i] = leaf->recs[nr_keep + i];
564 		xfs_iext_rec_clear(&leaf->recs[nr_keep + i]);
565 	}
566 
567 	if (cur->pos >= nr_keep) {
568 		cur->leaf = new;
569 		cur->pos -= nr_keep;
570 		*nr_entries = nr_move;
571 	} else {
572 		*nr_entries = nr_keep;
573 	}
574 done:
575 	if (leaf->next)
576 		leaf->next->prev = new;
577 	new->next = leaf->next;
578 	new->prev = leaf;
579 	leaf->next = new;
580 	return new;
581 }
582 
583 static void
584 xfs_iext_alloc_root(
585 	struct xfs_ifork	*ifp,
586 	struct xfs_iext_cursor	*cur)
587 {
588 	ASSERT(ifp->if_bytes == 0);
589 
590 	ifp->if_u1.if_root = kmem_zalloc(sizeof(struct xfs_iext_rec), KM_NOFS);
591 	ifp->if_height = 1;
592 
593 	/* now that we have a node step into it */
594 	cur->leaf = ifp->if_u1.if_root;
595 	cur->pos = 0;
596 }
597 
598 static void
599 xfs_iext_realloc_root(
600 	struct xfs_ifork	*ifp,
601 	struct xfs_iext_cursor	*cur)
602 {
603 	size_t new_size = ifp->if_bytes + sizeof(struct xfs_iext_rec);
604 	void *new;
605 
606 	/* account for the prev/next pointers */
607 	if (new_size / sizeof(struct xfs_iext_rec) == RECS_PER_LEAF)
608 		new_size = NODE_SIZE;
609 
610 	new = kmem_realloc(ifp->if_u1.if_root, new_size, KM_NOFS);
611 	memset(new + ifp->if_bytes, 0, new_size - ifp->if_bytes);
612 	ifp->if_u1.if_root = new;
613 	cur->leaf = new;
614 }
615 
616 /*
617  * Increment the sequence counter on extent tree changes. If we are on a COW
618  * fork, this allows the writeback code to skip looking for a COW extent if the
619  * COW fork hasn't changed. We use WRITE_ONCE here to ensure the update to the
620  * sequence counter is seen before the modifications to the extent tree itself
621  * take effect.
622  */
623 static inline void xfs_iext_inc_seq(struct xfs_ifork *ifp, int state)
624 {
625 	WRITE_ONCE(ifp->if_seq, READ_ONCE(ifp->if_seq) + 1);
626 }
627 
628 void
629 xfs_iext_insert(
630 	struct xfs_inode	*ip,
631 	struct xfs_iext_cursor	*cur,
632 	struct xfs_bmbt_irec	*irec,
633 	int			state)
634 {
635 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
636 	xfs_fileoff_t		offset = irec->br_startoff;
637 	struct xfs_iext_leaf	*new = NULL;
638 	int			nr_entries, i;
639 
640 	xfs_iext_inc_seq(ifp, state);
641 
642 	if (ifp->if_height == 0)
643 		xfs_iext_alloc_root(ifp, cur);
644 	else if (ifp->if_height == 1)
645 		xfs_iext_realloc_root(ifp, cur);
646 
647 	nr_entries = xfs_iext_leaf_nr_entries(ifp, cur->leaf, cur->pos);
648 	ASSERT(nr_entries <= RECS_PER_LEAF);
649 	ASSERT(cur->pos >= nr_entries ||
650 	       xfs_iext_rec_cmp(cur_rec(cur), irec->br_startoff) != 0);
651 
652 	if (nr_entries == RECS_PER_LEAF)
653 		new = xfs_iext_split_leaf(cur, &nr_entries);
654 
655 	/*
656 	 * Update the pointers in higher levels if the first entry changes
657 	 * in an existing node.
658 	 */
659 	if (cur->leaf != new && cur->pos == 0 && nr_entries > 0) {
660 		xfs_iext_update_node(ifp, xfs_iext_leaf_key(cur->leaf, 0),
661 				offset, 1, cur->leaf);
662 	}
663 
664 	for (i = nr_entries; i > cur->pos; i--)
665 		cur->leaf->recs[i] = cur->leaf->recs[i - 1];
666 	xfs_iext_set(cur_rec(cur), irec);
667 	ifp->if_bytes += sizeof(struct xfs_iext_rec);
668 
669 	trace_xfs_iext_insert(ip, cur, state, _RET_IP_);
670 
671 	if (new)
672 		xfs_iext_insert_node(ifp, xfs_iext_leaf_key(new, 0), new, 2);
673 }
674 
675 static struct xfs_iext_node *
676 xfs_iext_rebalance_node(
677 	struct xfs_iext_node	*parent,
678 	int			*pos,
679 	struct xfs_iext_node	*node,
680 	int			nr_entries)
681 {
682 	/*
683 	 * If the neighbouring nodes are completely full, or have different
684 	 * parents, we might never be able to merge our node, and will only
685 	 * delete it once the number of entries hits zero.
686 	 */
687 	if (nr_entries == 0)
688 		return node;
689 
690 	if (*pos > 0) {
691 		struct xfs_iext_node *prev = parent->ptrs[*pos - 1];
692 		int nr_prev = xfs_iext_node_nr_entries(prev, 0), i;
693 
694 		if (nr_prev + nr_entries <= KEYS_PER_NODE) {
695 			for (i = 0; i < nr_entries; i++) {
696 				prev->keys[nr_prev + i] = node->keys[i];
697 				prev->ptrs[nr_prev + i] = node->ptrs[i];
698 			}
699 			return node;
700 		}
701 	}
702 
703 	if (*pos + 1 < xfs_iext_node_nr_entries(parent, *pos)) {
704 		struct xfs_iext_node *next = parent->ptrs[*pos + 1];
705 		int nr_next = xfs_iext_node_nr_entries(next, 0), i;
706 
707 		if (nr_entries + nr_next <= KEYS_PER_NODE) {
708 			/*
709 			 * Merge the next node into this node so that we don't
710 			 * have to do an additional update of the keys in the
711 			 * higher levels.
712 			 */
713 			for (i = 0; i < nr_next; i++) {
714 				node->keys[nr_entries + i] = next->keys[i];
715 				node->ptrs[nr_entries + i] = next->ptrs[i];
716 			}
717 
718 			++*pos;
719 			return next;
720 		}
721 	}
722 
723 	return NULL;
724 }
725 
726 static void
727 xfs_iext_remove_node(
728 	struct xfs_ifork	*ifp,
729 	xfs_fileoff_t		offset,
730 	void			*victim)
731 {
732 	struct xfs_iext_node	*node, *parent;
733 	int			level = 2, pos, nr_entries, i;
734 
735 	ASSERT(level <= ifp->if_height);
736 	node = xfs_iext_find_level(ifp, offset, level);
737 	pos = xfs_iext_node_pos(node, offset);
738 again:
739 	ASSERT(node->ptrs[pos]);
740 	ASSERT(node->ptrs[pos] == victim);
741 	kmem_free(victim);
742 
743 	nr_entries = xfs_iext_node_nr_entries(node, pos) - 1;
744 	offset = node->keys[0];
745 	for (i = pos; i < nr_entries; i++) {
746 		node->keys[i] = node->keys[i + 1];
747 		node->ptrs[i] = node->ptrs[i + 1];
748 	}
749 	node->keys[nr_entries] = XFS_IEXT_KEY_INVALID;
750 	node->ptrs[nr_entries] = NULL;
751 
752 	if (pos == 0 && nr_entries > 0) {
753 		xfs_iext_update_node(ifp, offset, node->keys[0], level, node);
754 		offset = node->keys[0];
755 	}
756 
757 	if (nr_entries >= KEYS_PER_NODE / 2)
758 		return;
759 
760 	if (level < ifp->if_height) {
761 		/*
762 		 * If we aren't at the root yet try to find a neighbour node to
763 		 * merge with (or delete the node if it is empty), and then
764 		 * recurse up to the next level.
765 		 */
766 		level++;
767 		parent = xfs_iext_find_level(ifp, offset, level);
768 		pos = xfs_iext_node_pos(parent, offset);
769 
770 		ASSERT(pos != KEYS_PER_NODE);
771 		ASSERT(parent->ptrs[pos] == node);
772 
773 		node = xfs_iext_rebalance_node(parent, &pos, node, nr_entries);
774 		if (node) {
775 			victim = node;
776 			node = parent;
777 			goto again;
778 		}
779 	} else if (nr_entries == 1) {
780 		/*
781 		 * If we are at the root and only one entry is left we can just
782 		 * free this node and update the root pointer.
783 		 */
784 		ASSERT(node == ifp->if_u1.if_root);
785 		ifp->if_u1.if_root = node->ptrs[0];
786 		ifp->if_height--;
787 		kmem_free(node);
788 	}
789 }
790 
791 static void
792 xfs_iext_rebalance_leaf(
793 	struct xfs_ifork	*ifp,
794 	struct xfs_iext_cursor	*cur,
795 	struct xfs_iext_leaf	*leaf,
796 	xfs_fileoff_t		offset,
797 	int			nr_entries)
798 {
799 	/*
800 	 * If the neighbouring nodes are completely full we might never be able
801 	 * to merge our node, and will only delete it once the number of
802 	 * entries hits zero.
803 	 */
804 	if (nr_entries == 0)
805 		goto remove_node;
806 
807 	if (leaf->prev) {
808 		int nr_prev = xfs_iext_leaf_nr_entries(ifp, leaf->prev, 0), i;
809 
810 		if (nr_prev + nr_entries <= RECS_PER_LEAF) {
811 			for (i = 0; i < nr_entries; i++)
812 				leaf->prev->recs[nr_prev + i] = leaf->recs[i];
813 
814 			if (cur->leaf == leaf) {
815 				cur->leaf = leaf->prev;
816 				cur->pos += nr_prev;
817 			}
818 			goto remove_node;
819 		}
820 	}
821 
822 	if (leaf->next) {
823 		int nr_next = xfs_iext_leaf_nr_entries(ifp, leaf->next, 0), i;
824 
825 		if (nr_entries + nr_next <= RECS_PER_LEAF) {
826 			/*
827 			 * Merge the next node into this node so that we don't
828 			 * have to do an additional update of the keys in the
829 			 * higher levels.
830 			 */
831 			for (i = 0; i < nr_next; i++) {
832 				leaf->recs[nr_entries + i] =
833 					leaf->next->recs[i];
834 			}
835 
836 			if (cur->leaf == leaf->next) {
837 				cur->leaf = leaf;
838 				cur->pos += nr_entries;
839 			}
840 
841 			offset = xfs_iext_leaf_key(leaf->next, 0);
842 			leaf = leaf->next;
843 			goto remove_node;
844 		}
845 	}
846 
847 	return;
848 remove_node:
849 	if (leaf->prev)
850 		leaf->prev->next = leaf->next;
851 	if (leaf->next)
852 		leaf->next->prev = leaf->prev;
853 	xfs_iext_remove_node(ifp, offset, leaf);
854 }
855 
856 static void
857 xfs_iext_free_last_leaf(
858 	struct xfs_ifork	*ifp)
859 {
860 	ifp->if_height--;
861 	kmem_free(ifp->if_u1.if_root);
862 	ifp->if_u1.if_root = NULL;
863 }
864 
865 void
866 xfs_iext_remove(
867 	struct xfs_inode	*ip,
868 	struct xfs_iext_cursor	*cur,
869 	int			state)
870 {
871 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
872 	struct xfs_iext_leaf	*leaf = cur->leaf;
873 	xfs_fileoff_t		offset = xfs_iext_leaf_key(leaf, 0);
874 	int			i, nr_entries;
875 
876 	trace_xfs_iext_remove(ip, cur, state, _RET_IP_);
877 
878 	ASSERT(ifp->if_height > 0);
879 	ASSERT(ifp->if_u1.if_root != NULL);
880 	ASSERT(xfs_iext_valid(ifp, cur));
881 
882 	xfs_iext_inc_seq(ifp, state);
883 
884 	nr_entries = xfs_iext_leaf_nr_entries(ifp, leaf, cur->pos) - 1;
885 	for (i = cur->pos; i < nr_entries; i++)
886 		leaf->recs[i] = leaf->recs[i + 1];
887 	xfs_iext_rec_clear(&leaf->recs[nr_entries]);
888 	ifp->if_bytes -= sizeof(struct xfs_iext_rec);
889 
890 	if (cur->pos == 0 && nr_entries > 0) {
891 		xfs_iext_update_node(ifp, offset, xfs_iext_leaf_key(leaf, 0), 1,
892 				leaf);
893 		offset = xfs_iext_leaf_key(leaf, 0);
894 	} else if (cur->pos == nr_entries) {
895 		if (ifp->if_height > 1 && leaf->next)
896 			cur->leaf = leaf->next;
897 		else
898 			cur->leaf = NULL;
899 		cur->pos = 0;
900 	}
901 
902 	if (nr_entries >= RECS_PER_LEAF / 2)
903 		return;
904 
905 	if (ifp->if_height > 1)
906 		xfs_iext_rebalance_leaf(ifp, cur, leaf, offset, nr_entries);
907 	else if (nr_entries == 0)
908 		xfs_iext_free_last_leaf(ifp);
909 }
910 
911 /*
912  * Lookup the extent covering bno.
913  *
914  * If there is an extent covering bno return the extent index, and store the
915  * expanded extent structure in *gotp, and the extent cursor in *cur.
916  * If there is no extent covering bno, but there is an extent after it (e.g.
917  * it lies in a hole) return that extent in *gotp and its cursor in *cur
918  * instead.
919  * If bno is beyond the last extent return false, and return an invalid
920  * cursor value.
921  */
922 bool
923 xfs_iext_lookup_extent(
924 	struct xfs_inode	*ip,
925 	struct xfs_ifork	*ifp,
926 	xfs_fileoff_t		offset,
927 	struct xfs_iext_cursor	*cur,
928 	struct xfs_bmbt_irec	*gotp)
929 {
930 	XFS_STATS_INC(ip->i_mount, xs_look_exlist);
931 
932 	cur->leaf = xfs_iext_find_level(ifp, offset, 1);
933 	if (!cur->leaf) {
934 		cur->pos = 0;
935 		return false;
936 	}
937 
938 	for (cur->pos = 0; cur->pos < xfs_iext_max_recs(ifp); cur->pos++) {
939 		struct xfs_iext_rec *rec = cur_rec(cur);
940 
941 		if (xfs_iext_rec_is_empty(rec))
942 			break;
943 		if (xfs_iext_rec_cmp(rec, offset) >= 0)
944 			goto found;
945 	}
946 
947 	/* Try looking in the next node for an entry > offset */
948 	if (ifp->if_height == 1 || !cur->leaf->next)
949 		return false;
950 	cur->leaf = cur->leaf->next;
951 	cur->pos = 0;
952 	if (!xfs_iext_valid(ifp, cur))
953 		return false;
954 found:
955 	xfs_iext_get(gotp, cur_rec(cur));
956 	return true;
957 }
958 
959 /*
960  * Returns the last extent before end, and if this extent doesn't cover
961  * end, update end to the end of the extent.
962  */
963 bool
964 xfs_iext_lookup_extent_before(
965 	struct xfs_inode	*ip,
966 	struct xfs_ifork	*ifp,
967 	xfs_fileoff_t		*end,
968 	struct xfs_iext_cursor	*cur,
969 	struct xfs_bmbt_irec	*gotp)
970 {
971 	/* could be optimized to not even look up the next on a match.. */
972 	if (xfs_iext_lookup_extent(ip, ifp, *end - 1, cur, gotp) &&
973 	    gotp->br_startoff <= *end - 1)
974 		return true;
975 	if (!xfs_iext_prev_extent(ifp, cur, gotp))
976 		return false;
977 	*end = gotp->br_startoff + gotp->br_blockcount;
978 	return true;
979 }
980 
981 void
982 xfs_iext_update_extent(
983 	struct xfs_inode	*ip,
984 	int			state,
985 	struct xfs_iext_cursor	*cur,
986 	struct xfs_bmbt_irec	*new)
987 {
988 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
989 
990 	xfs_iext_inc_seq(ifp, state);
991 
992 	if (cur->pos == 0) {
993 		struct xfs_bmbt_irec	old;
994 
995 		xfs_iext_get(&old, cur_rec(cur));
996 		if (new->br_startoff != old.br_startoff) {
997 			xfs_iext_update_node(ifp, old.br_startoff,
998 					new->br_startoff, 1, cur->leaf);
999 		}
1000 	}
1001 
1002 	trace_xfs_bmap_pre_update(ip, cur, state, _RET_IP_);
1003 	xfs_iext_set(cur_rec(cur), new);
1004 	trace_xfs_bmap_post_update(ip, cur, state, _RET_IP_);
1005 }
1006 
1007 /*
1008  * Return true if the cursor points at an extent and return the extent structure
1009  * in gotp.  Else return false.
1010  */
1011 bool
1012 xfs_iext_get_extent(
1013 	struct xfs_ifork	*ifp,
1014 	struct xfs_iext_cursor	*cur,
1015 	struct xfs_bmbt_irec	*gotp)
1016 {
1017 	if (!xfs_iext_valid(ifp, cur))
1018 		return false;
1019 	xfs_iext_get(gotp, cur_rec(cur));
1020 	return true;
1021 }
1022 
1023 /*
1024  * This is a recursive function, because of that we need to be extremely
1025  * careful with stack usage.
1026  */
1027 static void
1028 xfs_iext_destroy_node(
1029 	struct xfs_iext_node	*node,
1030 	int			level)
1031 {
1032 	int			i;
1033 
1034 	if (level > 1) {
1035 		for (i = 0; i < KEYS_PER_NODE; i++) {
1036 			if (node->keys[i] == XFS_IEXT_KEY_INVALID)
1037 				break;
1038 			xfs_iext_destroy_node(node->ptrs[i], level - 1);
1039 		}
1040 	}
1041 
1042 	kmem_free(node);
1043 }
1044 
1045 void
1046 xfs_iext_destroy(
1047 	struct xfs_ifork	*ifp)
1048 {
1049 	xfs_iext_destroy_node(ifp->if_u1.if_root, ifp->if_height);
1050 
1051 	ifp->if_bytes = 0;
1052 	ifp->if_height = 0;
1053 	ifp->if_u1.if_root = NULL;
1054 }
1055