xref: /linux/fs/xfs/libxfs/xfs_iext_tree.c (revision 17cfcb68af3bc7d5e8ae08779b1853310a2949f3)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2017 Christoph Hellwig.
4  */
5 
6 #include "xfs.h"
7 #include "xfs_shared.h"
8 #include "xfs_format.h"
9 #include "xfs_bit.h"
10 #include "xfs_log_format.h"
11 #include "xfs_inode.h"
12 #include "xfs_trans_resv.h"
13 #include "xfs_mount.h"
14 #include "xfs_trace.h"
15 
16 /*
17  * In-core extent record layout:
18  *
19  * +-------+----------------------------+
20  * | 00:53 | all 54 bits of startoff    |
21  * | 54:63 | low 10 bits of startblock  |
22  * +-------+----------------------------+
23  * | 00:20 | all 21 bits of length      |
24  * |    21 | unwritten extent bit       |
25  * | 22:63 | high 42 bits of startblock |
26  * +-------+----------------------------+
27  */
28 #define XFS_IEXT_STARTOFF_MASK		xfs_mask64lo(BMBT_STARTOFF_BITLEN)
29 #define XFS_IEXT_LENGTH_MASK		xfs_mask64lo(BMBT_BLOCKCOUNT_BITLEN)
30 #define XFS_IEXT_STARTBLOCK_MASK	xfs_mask64lo(BMBT_STARTBLOCK_BITLEN)
31 
32 struct xfs_iext_rec {
33 	uint64_t			lo;
34 	uint64_t			hi;
35 };
36 
37 /*
38  * Given that the length can't be a zero, only an empty hi value indicates an
39  * unused record.
40  */
41 static bool xfs_iext_rec_is_empty(struct xfs_iext_rec *rec)
42 {
43 	return rec->hi == 0;
44 }
45 
46 static inline void xfs_iext_rec_clear(struct xfs_iext_rec *rec)
47 {
48 	rec->lo = 0;
49 	rec->hi = 0;
50 }
51 
52 static void
53 xfs_iext_set(
54 	struct xfs_iext_rec	*rec,
55 	struct xfs_bmbt_irec	*irec)
56 {
57 	ASSERT((irec->br_startoff & ~XFS_IEXT_STARTOFF_MASK) == 0);
58 	ASSERT((irec->br_blockcount & ~XFS_IEXT_LENGTH_MASK) == 0);
59 	ASSERT((irec->br_startblock & ~XFS_IEXT_STARTBLOCK_MASK) == 0);
60 
61 	rec->lo = irec->br_startoff & XFS_IEXT_STARTOFF_MASK;
62 	rec->hi = irec->br_blockcount & XFS_IEXT_LENGTH_MASK;
63 
64 	rec->lo |= (irec->br_startblock << 54);
65 	rec->hi |= ((irec->br_startblock & ~xfs_mask64lo(10)) << (22 - 10));
66 
67 	if (irec->br_state == XFS_EXT_UNWRITTEN)
68 		rec->hi |= (1 << 21);
69 }
70 
71 static void
72 xfs_iext_get(
73 	struct xfs_bmbt_irec	*irec,
74 	struct xfs_iext_rec	*rec)
75 {
76 	irec->br_startoff = rec->lo & XFS_IEXT_STARTOFF_MASK;
77 	irec->br_blockcount = rec->hi & XFS_IEXT_LENGTH_MASK;
78 
79 	irec->br_startblock = rec->lo >> 54;
80 	irec->br_startblock |= (rec->hi & xfs_mask64hi(42)) >> (22 - 10);
81 
82 	if (rec->hi & (1 << 21))
83 		irec->br_state = XFS_EXT_UNWRITTEN;
84 	else
85 		irec->br_state = XFS_EXT_NORM;
86 }
87 
88 enum {
89 	NODE_SIZE	= 256,
90 	KEYS_PER_NODE	= NODE_SIZE / (sizeof(uint64_t) + sizeof(void *)),
91 	RECS_PER_LEAF	= (NODE_SIZE - (2 * sizeof(struct xfs_iext_leaf *))) /
92 				sizeof(struct xfs_iext_rec),
93 };
94 
95 /*
96  * In-core extent btree block layout:
97  *
98  * There are two types of blocks in the btree: leaf and inner (non-leaf) blocks.
99  *
100  * The leaf blocks are made up by %KEYS_PER_NODE extent records, which each
101  * contain the startoffset, blockcount, startblock and unwritten extent flag.
102  * See above for the exact format, followed by pointers to the previous and next
103  * leaf blocks (if there are any).
104  *
105  * The inner (non-leaf) blocks first contain KEYS_PER_NODE lookup keys, followed
106  * by an equal number of pointers to the btree blocks at the next lower level.
107  *
108  *		+-------+-------+-------+-------+-------+----------+----------+
109  * Leaf:	| rec 1 | rec 2 | rec 3 | rec 4 | rec N | prev-ptr | next-ptr |
110  *		+-------+-------+-------+-------+-------+----------+----------+
111  *
112  *		+-------+-------+-------+-------+-------+-------+------+-------+
113  * Inner:	| key 1 | key 2 | key 3 | key N | ptr 1 | ptr 2 | ptr3 | ptr N |
114  *		+-------+-------+-------+-------+-------+-------+------+-------+
115  */
116 struct xfs_iext_node {
117 	uint64_t		keys[KEYS_PER_NODE];
118 #define XFS_IEXT_KEY_INVALID	(1ULL << 63)
119 	void			*ptrs[KEYS_PER_NODE];
120 };
121 
122 struct xfs_iext_leaf {
123 	struct xfs_iext_rec	recs[RECS_PER_LEAF];
124 	struct xfs_iext_leaf	*prev;
125 	struct xfs_iext_leaf	*next;
126 };
127 
128 inline xfs_extnum_t xfs_iext_count(struct xfs_ifork *ifp)
129 {
130 	return ifp->if_bytes / sizeof(struct xfs_iext_rec);
131 }
132 
133 static inline int xfs_iext_max_recs(struct xfs_ifork *ifp)
134 {
135 	if (ifp->if_height == 1)
136 		return xfs_iext_count(ifp);
137 	return RECS_PER_LEAF;
138 }
139 
140 static inline struct xfs_iext_rec *cur_rec(struct xfs_iext_cursor *cur)
141 {
142 	return &cur->leaf->recs[cur->pos];
143 }
144 
145 static inline bool xfs_iext_valid(struct xfs_ifork *ifp,
146 		struct xfs_iext_cursor *cur)
147 {
148 	if (!cur->leaf)
149 		return false;
150 	if (cur->pos < 0 || cur->pos >= xfs_iext_max_recs(ifp))
151 		return false;
152 	if (xfs_iext_rec_is_empty(cur_rec(cur)))
153 		return false;
154 	return true;
155 }
156 
157 static void *
158 xfs_iext_find_first_leaf(
159 	struct xfs_ifork	*ifp)
160 {
161 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
162 	int			height;
163 
164 	if (!ifp->if_height)
165 		return NULL;
166 
167 	for (height = ifp->if_height; height > 1; height--) {
168 		node = node->ptrs[0];
169 		ASSERT(node);
170 	}
171 
172 	return node;
173 }
174 
175 static void *
176 xfs_iext_find_last_leaf(
177 	struct xfs_ifork	*ifp)
178 {
179 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
180 	int			height, i;
181 
182 	if (!ifp->if_height)
183 		return NULL;
184 
185 	for (height = ifp->if_height; height > 1; height--) {
186 		for (i = 1; i < KEYS_PER_NODE; i++)
187 			if (!node->ptrs[i])
188 				break;
189 		node = node->ptrs[i - 1];
190 		ASSERT(node);
191 	}
192 
193 	return node;
194 }
195 
196 void
197 xfs_iext_first(
198 	struct xfs_ifork	*ifp,
199 	struct xfs_iext_cursor	*cur)
200 {
201 	cur->pos = 0;
202 	cur->leaf = xfs_iext_find_first_leaf(ifp);
203 }
204 
205 void
206 xfs_iext_last(
207 	struct xfs_ifork	*ifp,
208 	struct xfs_iext_cursor	*cur)
209 {
210 	int			i;
211 
212 	cur->leaf = xfs_iext_find_last_leaf(ifp);
213 	if (!cur->leaf) {
214 		cur->pos = 0;
215 		return;
216 	}
217 
218 	for (i = 1; i < xfs_iext_max_recs(ifp); i++) {
219 		if (xfs_iext_rec_is_empty(&cur->leaf->recs[i]))
220 			break;
221 	}
222 	cur->pos = i - 1;
223 }
224 
225 void
226 xfs_iext_next(
227 	struct xfs_ifork	*ifp,
228 	struct xfs_iext_cursor	*cur)
229 {
230 	if (!cur->leaf) {
231 		ASSERT(cur->pos <= 0 || cur->pos >= RECS_PER_LEAF);
232 		xfs_iext_first(ifp, cur);
233 		return;
234 	}
235 
236 	ASSERT(cur->pos >= 0);
237 	ASSERT(cur->pos < xfs_iext_max_recs(ifp));
238 
239 	cur->pos++;
240 	if (ifp->if_height > 1 && !xfs_iext_valid(ifp, cur) &&
241 	    cur->leaf->next) {
242 		cur->leaf = cur->leaf->next;
243 		cur->pos = 0;
244 	}
245 }
246 
247 void
248 xfs_iext_prev(
249 	struct xfs_ifork	*ifp,
250 	struct xfs_iext_cursor	*cur)
251 {
252 	if (!cur->leaf) {
253 		ASSERT(cur->pos <= 0 || cur->pos >= RECS_PER_LEAF);
254 		xfs_iext_last(ifp, cur);
255 		return;
256 	}
257 
258 	ASSERT(cur->pos >= 0);
259 	ASSERT(cur->pos <= RECS_PER_LEAF);
260 
261 recurse:
262 	do {
263 		cur->pos--;
264 		if (xfs_iext_valid(ifp, cur))
265 			return;
266 	} while (cur->pos > 0);
267 
268 	if (ifp->if_height > 1 && cur->leaf->prev) {
269 		cur->leaf = cur->leaf->prev;
270 		cur->pos = RECS_PER_LEAF;
271 		goto recurse;
272 	}
273 }
274 
275 static inline int
276 xfs_iext_key_cmp(
277 	struct xfs_iext_node	*node,
278 	int			n,
279 	xfs_fileoff_t		offset)
280 {
281 	if (node->keys[n] > offset)
282 		return 1;
283 	if (node->keys[n] < offset)
284 		return -1;
285 	return 0;
286 }
287 
288 static inline int
289 xfs_iext_rec_cmp(
290 	struct xfs_iext_rec	*rec,
291 	xfs_fileoff_t		offset)
292 {
293 	uint64_t		rec_offset = rec->lo & XFS_IEXT_STARTOFF_MASK;
294 	uint32_t		rec_len = rec->hi & XFS_IEXT_LENGTH_MASK;
295 
296 	if (rec_offset > offset)
297 		return 1;
298 	if (rec_offset + rec_len <= offset)
299 		return -1;
300 	return 0;
301 }
302 
303 static void *
304 xfs_iext_find_level(
305 	struct xfs_ifork	*ifp,
306 	xfs_fileoff_t		offset,
307 	int			level)
308 {
309 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
310 	int			height, i;
311 
312 	if (!ifp->if_height)
313 		return NULL;
314 
315 	for (height = ifp->if_height; height > level; height--) {
316 		for (i = 1; i < KEYS_PER_NODE; i++)
317 			if (xfs_iext_key_cmp(node, i, offset) > 0)
318 				break;
319 
320 		node = node->ptrs[i - 1];
321 		if (!node)
322 			break;
323 	}
324 
325 	return node;
326 }
327 
328 static int
329 xfs_iext_node_pos(
330 	struct xfs_iext_node	*node,
331 	xfs_fileoff_t		offset)
332 {
333 	int			i;
334 
335 	for (i = 1; i < KEYS_PER_NODE; i++) {
336 		if (xfs_iext_key_cmp(node, i, offset) > 0)
337 			break;
338 	}
339 
340 	return i - 1;
341 }
342 
343 static int
344 xfs_iext_node_insert_pos(
345 	struct xfs_iext_node	*node,
346 	xfs_fileoff_t		offset)
347 {
348 	int			i;
349 
350 	for (i = 0; i < KEYS_PER_NODE; i++) {
351 		if (xfs_iext_key_cmp(node, i, offset) > 0)
352 			return i;
353 	}
354 
355 	return KEYS_PER_NODE;
356 }
357 
358 static int
359 xfs_iext_node_nr_entries(
360 	struct xfs_iext_node	*node,
361 	int			start)
362 {
363 	int			i;
364 
365 	for (i = start; i < KEYS_PER_NODE; i++) {
366 		if (node->keys[i] == XFS_IEXT_KEY_INVALID)
367 			break;
368 	}
369 
370 	return i;
371 }
372 
373 static int
374 xfs_iext_leaf_nr_entries(
375 	struct xfs_ifork	*ifp,
376 	struct xfs_iext_leaf	*leaf,
377 	int			start)
378 {
379 	int			i;
380 
381 	for (i = start; i < xfs_iext_max_recs(ifp); i++) {
382 		if (xfs_iext_rec_is_empty(&leaf->recs[i]))
383 			break;
384 	}
385 
386 	return i;
387 }
388 
389 static inline uint64_t
390 xfs_iext_leaf_key(
391 	struct xfs_iext_leaf	*leaf,
392 	int			n)
393 {
394 	return leaf->recs[n].lo & XFS_IEXT_STARTOFF_MASK;
395 }
396 
397 static void
398 xfs_iext_grow(
399 	struct xfs_ifork	*ifp)
400 {
401 	struct xfs_iext_node	*node = kmem_zalloc(NODE_SIZE, KM_NOFS);
402 	int			i;
403 
404 	if (ifp->if_height == 1) {
405 		struct xfs_iext_leaf *prev = ifp->if_u1.if_root;
406 
407 		node->keys[0] = xfs_iext_leaf_key(prev, 0);
408 		node->ptrs[0] = prev;
409 	} else  {
410 		struct xfs_iext_node *prev = ifp->if_u1.if_root;
411 
412 		ASSERT(ifp->if_height > 1);
413 
414 		node->keys[0] = prev->keys[0];
415 		node->ptrs[0] = prev;
416 	}
417 
418 	for (i = 1; i < KEYS_PER_NODE; i++)
419 		node->keys[i] = XFS_IEXT_KEY_INVALID;
420 
421 	ifp->if_u1.if_root = node;
422 	ifp->if_height++;
423 }
424 
425 static void
426 xfs_iext_update_node(
427 	struct xfs_ifork	*ifp,
428 	xfs_fileoff_t		old_offset,
429 	xfs_fileoff_t		new_offset,
430 	int			level,
431 	void			*ptr)
432 {
433 	struct xfs_iext_node	*node = ifp->if_u1.if_root;
434 	int			height, i;
435 
436 	for (height = ifp->if_height; height > level; height--) {
437 		for (i = 0; i < KEYS_PER_NODE; i++) {
438 			if (i > 0 && xfs_iext_key_cmp(node, i, old_offset) > 0)
439 				break;
440 			if (node->keys[i] == old_offset)
441 				node->keys[i] = new_offset;
442 		}
443 		node = node->ptrs[i - 1];
444 		ASSERT(node);
445 	}
446 
447 	ASSERT(node == ptr);
448 }
449 
450 static struct xfs_iext_node *
451 xfs_iext_split_node(
452 	struct xfs_iext_node	**nodep,
453 	int			*pos,
454 	int			*nr_entries)
455 {
456 	struct xfs_iext_node	*node = *nodep;
457 	struct xfs_iext_node	*new = kmem_zalloc(NODE_SIZE, KM_NOFS);
458 	const int		nr_move = KEYS_PER_NODE / 2;
459 	int			nr_keep = nr_move + (KEYS_PER_NODE & 1);
460 	int			i = 0;
461 
462 	/* for sequential append operations just spill over into the new node */
463 	if (*pos == KEYS_PER_NODE) {
464 		*nodep = new;
465 		*pos = 0;
466 		*nr_entries = 0;
467 		goto done;
468 	}
469 
470 
471 	for (i = 0; i < nr_move; i++) {
472 		new->keys[i] = node->keys[nr_keep + i];
473 		new->ptrs[i] = node->ptrs[nr_keep + i];
474 
475 		node->keys[nr_keep + i] = XFS_IEXT_KEY_INVALID;
476 		node->ptrs[nr_keep + i] = NULL;
477 	}
478 
479 	if (*pos >= nr_keep) {
480 		*nodep = new;
481 		*pos -= nr_keep;
482 		*nr_entries = nr_move;
483 	} else {
484 		*nr_entries = nr_keep;
485 	}
486 done:
487 	for (; i < KEYS_PER_NODE; i++)
488 		new->keys[i] = XFS_IEXT_KEY_INVALID;
489 	return new;
490 }
491 
492 static void
493 xfs_iext_insert_node(
494 	struct xfs_ifork	*ifp,
495 	uint64_t		offset,
496 	void			*ptr,
497 	int			level)
498 {
499 	struct xfs_iext_node	*node, *new;
500 	int			i, pos, nr_entries;
501 
502 again:
503 	if (ifp->if_height < level)
504 		xfs_iext_grow(ifp);
505 
506 	new = NULL;
507 	node = xfs_iext_find_level(ifp, offset, level);
508 	pos = xfs_iext_node_insert_pos(node, offset);
509 	nr_entries = xfs_iext_node_nr_entries(node, pos);
510 
511 	ASSERT(pos >= nr_entries || xfs_iext_key_cmp(node, pos, offset) != 0);
512 	ASSERT(nr_entries <= KEYS_PER_NODE);
513 
514 	if (nr_entries == KEYS_PER_NODE)
515 		new = xfs_iext_split_node(&node, &pos, &nr_entries);
516 
517 	/*
518 	 * Update the pointers in higher levels if the first entry changes
519 	 * in an existing node.
520 	 */
521 	if (node != new && pos == 0 && nr_entries > 0)
522 		xfs_iext_update_node(ifp, node->keys[0], offset, level, node);
523 
524 	for (i = nr_entries; i > pos; i--) {
525 		node->keys[i] = node->keys[i - 1];
526 		node->ptrs[i] = node->ptrs[i - 1];
527 	}
528 	node->keys[pos] = offset;
529 	node->ptrs[pos] = ptr;
530 
531 	if (new) {
532 		offset = new->keys[0];
533 		ptr = new;
534 		level++;
535 		goto again;
536 	}
537 }
538 
539 static struct xfs_iext_leaf *
540 xfs_iext_split_leaf(
541 	struct xfs_iext_cursor	*cur,
542 	int			*nr_entries)
543 {
544 	struct xfs_iext_leaf	*leaf = cur->leaf;
545 	struct xfs_iext_leaf	*new = kmem_zalloc(NODE_SIZE, KM_NOFS);
546 	const int		nr_move = RECS_PER_LEAF / 2;
547 	int			nr_keep = nr_move + (RECS_PER_LEAF & 1);
548 	int			i;
549 
550 	/* for sequential append operations just spill over into the new node */
551 	if (cur->pos == RECS_PER_LEAF) {
552 		cur->leaf = new;
553 		cur->pos = 0;
554 		*nr_entries = 0;
555 		goto done;
556 	}
557 
558 	for (i = 0; i < nr_move; i++) {
559 		new->recs[i] = leaf->recs[nr_keep + i];
560 		xfs_iext_rec_clear(&leaf->recs[nr_keep + i]);
561 	}
562 
563 	if (cur->pos >= nr_keep) {
564 		cur->leaf = new;
565 		cur->pos -= nr_keep;
566 		*nr_entries = nr_move;
567 	} else {
568 		*nr_entries = nr_keep;
569 	}
570 done:
571 	if (leaf->next)
572 		leaf->next->prev = new;
573 	new->next = leaf->next;
574 	new->prev = leaf;
575 	leaf->next = new;
576 	return new;
577 }
578 
579 static void
580 xfs_iext_alloc_root(
581 	struct xfs_ifork	*ifp,
582 	struct xfs_iext_cursor	*cur)
583 {
584 	ASSERT(ifp->if_bytes == 0);
585 
586 	ifp->if_u1.if_root = kmem_zalloc(sizeof(struct xfs_iext_rec), KM_NOFS);
587 	ifp->if_height = 1;
588 
589 	/* now that we have a node step into it */
590 	cur->leaf = ifp->if_u1.if_root;
591 	cur->pos = 0;
592 }
593 
594 static void
595 xfs_iext_realloc_root(
596 	struct xfs_ifork	*ifp,
597 	struct xfs_iext_cursor	*cur)
598 {
599 	size_t new_size = ifp->if_bytes + sizeof(struct xfs_iext_rec);
600 	void *new;
601 
602 	/* account for the prev/next pointers */
603 	if (new_size / sizeof(struct xfs_iext_rec) == RECS_PER_LEAF)
604 		new_size = NODE_SIZE;
605 
606 	new = kmem_realloc(ifp->if_u1.if_root, new_size, KM_NOFS);
607 	memset(new + ifp->if_bytes, 0, new_size - ifp->if_bytes);
608 	ifp->if_u1.if_root = new;
609 	cur->leaf = new;
610 }
611 
612 /*
613  * Increment the sequence counter on extent tree changes. If we are on a COW
614  * fork, this allows the writeback code to skip looking for a COW extent if the
615  * COW fork hasn't changed. We use WRITE_ONCE here to ensure the update to the
616  * sequence counter is seen before the modifications to the extent tree itself
617  * take effect.
618  */
619 static inline void xfs_iext_inc_seq(struct xfs_ifork *ifp)
620 {
621 	WRITE_ONCE(ifp->if_seq, READ_ONCE(ifp->if_seq) + 1);
622 }
623 
624 void
625 xfs_iext_insert(
626 	struct xfs_inode	*ip,
627 	struct xfs_iext_cursor	*cur,
628 	struct xfs_bmbt_irec	*irec,
629 	int			state)
630 {
631 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
632 	xfs_fileoff_t		offset = irec->br_startoff;
633 	struct xfs_iext_leaf	*new = NULL;
634 	int			nr_entries, i;
635 
636 	xfs_iext_inc_seq(ifp);
637 
638 	if (ifp->if_height == 0)
639 		xfs_iext_alloc_root(ifp, cur);
640 	else if (ifp->if_height == 1)
641 		xfs_iext_realloc_root(ifp, cur);
642 
643 	nr_entries = xfs_iext_leaf_nr_entries(ifp, cur->leaf, cur->pos);
644 	ASSERT(nr_entries <= RECS_PER_LEAF);
645 	ASSERT(cur->pos >= nr_entries ||
646 	       xfs_iext_rec_cmp(cur_rec(cur), irec->br_startoff) != 0);
647 
648 	if (nr_entries == RECS_PER_LEAF)
649 		new = xfs_iext_split_leaf(cur, &nr_entries);
650 
651 	/*
652 	 * Update the pointers in higher levels if the first entry changes
653 	 * in an existing node.
654 	 */
655 	if (cur->leaf != new && cur->pos == 0 && nr_entries > 0) {
656 		xfs_iext_update_node(ifp, xfs_iext_leaf_key(cur->leaf, 0),
657 				offset, 1, cur->leaf);
658 	}
659 
660 	for (i = nr_entries; i > cur->pos; i--)
661 		cur->leaf->recs[i] = cur->leaf->recs[i - 1];
662 	xfs_iext_set(cur_rec(cur), irec);
663 	ifp->if_bytes += sizeof(struct xfs_iext_rec);
664 
665 	trace_xfs_iext_insert(ip, cur, state, _RET_IP_);
666 
667 	if (new)
668 		xfs_iext_insert_node(ifp, xfs_iext_leaf_key(new, 0), new, 2);
669 }
670 
671 static struct xfs_iext_node *
672 xfs_iext_rebalance_node(
673 	struct xfs_iext_node	*parent,
674 	int			*pos,
675 	struct xfs_iext_node	*node,
676 	int			nr_entries)
677 {
678 	/*
679 	 * If the neighbouring nodes are completely full, or have different
680 	 * parents, we might never be able to merge our node, and will only
681 	 * delete it once the number of entries hits zero.
682 	 */
683 	if (nr_entries == 0)
684 		return node;
685 
686 	if (*pos > 0) {
687 		struct xfs_iext_node *prev = parent->ptrs[*pos - 1];
688 		int nr_prev = xfs_iext_node_nr_entries(prev, 0), i;
689 
690 		if (nr_prev + nr_entries <= KEYS_PER_NODE) {
691 			for (i = 0; i < nr_entries; i++) {
692 				prev->keys[nr_prev + i] = node->keys[i];
693 				prev->ptrs[nr_prev + i] = node->ptrs[i];
694 			}
695 			return node;
696 		}
697 	}
698 
699 	if (*pos + 1 < xfs_iext_node_nr_entries(parent, *pos)) {
700 		struct xfs_iext_node *next = parent->ptrs[*pos + 1];
701 		int nr_next = xfs_iext_node_nr_entries(next, 0), i;
702 
703 		if (nr_entries + nr_next <= KEYS_PER_NODE) {
704 			/*
705 			 * Merge the next node into this node so that we don't
706 			 * have to do an additional update of the keys in the
707 			 * higher levels.
708 			 */
709 			for (i = 0; i < nr_next; i++) {
710 				node->keys[nr_entries + i] = next->keys[i];
711 				node->ptrs[nr_entries + i] = next->ptrs[i];
712 			}
713 
714 			++*pos;
715 			return next;
716 		}
717 	}
718 
719 	return NULL;
720 }
721 
722 static void
723 xfs_iext_remove_node(
724 	struct xfs_ifork	*ifp,
725 	xfs_fileoff_t		offset,
726 	void			*victim)
727 {
728 	struct xfs_iext_node	*node, *parent;
729 	int			level = 2, pos, nr_entries, i;
730 
731 	ASSERT(level <= ifp->if_height);
732 	node = xfs_iext_find_level(ifp, offset, level);
733 	pos = xfs_iext_node_pos(node, offset);
734 again:
735 	ASSERT(node->ptrs[pos]);
736 	ASSERT(node->ptrs[pos] == victim);
737 	kmem_free(victim);
738 
739 	nr_entries = xfs_iext_node_nr_entries(node, pos) - 1;
740 	offset = node->keys[0];
741 	for (i = pos; i < nr_entries; i++) {
742 		node->keys[i] = node->keys[i + 1];
743 		node->ptrs[i] = node->ptrs[i + 1];
744 	}
745 	node->keys[nr_entries] = XFS_IEXT_KEY_INVALID;
746 	node->ptrs[nr_entries] = NULL;
747 
748 	if (pos == 0 && nr_entries > 0) {
749 		xfs_iext_update_node(ifp, offset, node->keys[0], level, node);
750 		offset = node->keys[0];
751 	}
752 
753 	if (nr_entries >= KEYS_PER_NODE / 2)
754 		return;
755 
756 	if (level < ifp->if_height) {
757 		/*
758 		 * If we aren't at the root yet try to find a neighbour node to
759 		 * merge with (or delete the node if it is empty), and then
760 		 * recurse up to the next level.
761 		 */
762 		level++;
763 		parent = xfs_iext_find_level(ifp, offset, level);
764 		pos = xfs_iext_node_pos(parent, offset);
765 
766 		ASSERT(pos != KEYS_PER_NODE);
767 		ASSERT(parent->ptrs[pos] == node);
768 
769 		node = xfs_iext_rebalance_node(parent, &pos, node, nr_entries);
770 		if (node) {
771 			victim = node;
772 			node = parent;
773 			goto again;
774 		}
775 	} else if (nr_entries == 1) {
776 		/*
777 		 * If we are at the root and only one entry is left we can just
778 		 * free this node and update the root pointer.
779 		 */
780 		ASSERT(node == ifp->if_u1.if_root);
781 		ifp->if_u1.if_root = node->ptrs[0];
782 		ifp->if_height--;
783 		kmem_free(node);
784 	}
785 }
786 
787 static void
788 xfs_iext_rebalance_leaf(
789 	struct xfs_ifork	*ifp,
790 	struct xfs_iext_cursor	*cur,
791 	struct xfs_iext_leaf	*leaf,
792 	xfs_fileoff_t		offset,
793 	int			nr_entries)
794 {
795 	/*
796 	 * If the neighbouring nodes are completely full we might never be able
797 	 * to merge our node, and will only delete it once the number of
798 	 * entries hits zero.
799 	 */
800 	if (nr_entries == 0)
801 		goto remove_node;
802 
803 	if (leaf->prev) {
804 		int nr_prev = xfs_iext_leaf_nr_entries(ifp, leaf->prev, 0), i;
805 
806 		if (nr_prev + nr_entries <= RECS_PER_LEAF) {
807 			for (i = 0; i < nr_entries; i++)
808 				leaf->prev->recs[nr_prev + i] = leaf->recs[i];
809 
810 			if (cur->leaf == leaf) {
811 				cur->leaf = leaf->prev;
812 				cur->pos += nr_prev;
813 			}
814 			goto remove_node;
815 		}
816 	}
817 
818 	if (leaf->next) {
819 		int nr_next = xfs_iext_leaf_nr_entries(ifp, leaf->next, 0), i;
820 
821 		if (nr_entries + nr_next <= RECS_PER_LEAF) {
822 			/*
823 			 * Merge the next node into this node so that we don't
824 			 * have to do an additional update of the keys in the
825 			 * higher levels.
826 			 */
827 			for (i = 0; i < nr_next; i++) {
828 				leaf->recs[nr_entries + i] =
829 					leaf->next->recs[i];
830 			}
831 
832 			if (cur->leaf == leaf->next) {
833 				cur->leaf = leaf;
834 				cur->pos += nr_entries;
835 			}
836 
837 			offset = xfs_iext_leaf_key(leaf->next, 0);
838 			leaf = leaf->next;
839 			goto remove_node;
840 		}
841 	}
842 
843 	return;
844 remove_node:
845 	if (leaf->prev)
846 		leaf->prev->next = leaf->next;
847 	if (leaf->next)
848 		leaf->next->prev = leaf->prev;
849 	xfs_iext_remove_node(ifp, offset, leaf);
850 }
851 
852 static void
853 xfs_iext_free_last_leaf(
854 	struct xfs_ifork	*ifp)
855 {
856 	ifp->if_height--;
857 	kmem_free(ifp->if_u1.if_root);
858 	ifp->if_u1.if_root = NULL;
859 }
860 
861 void
862 xfs_iext_remove(
863 	struct xfs_inode	*ip,
864 	struct xfs_iext_cursor	*cur,
865 	int			state)
866 {
867 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
868 	struct xfs_iext_leaf	*leaf = cur->leaf;
869 	xfs_fileoff_t		offset = xfs_iext_leaf_key(leaf, 0);
870 	int			i, nr_entries;
871 
872 	trace_xfs_iext_remove(ip, cur, state, _RET_IP_);
873 
874 	ASSERT(ifp->if_height > 0);
875 	ASSERT(ifp->if_u1.if_root != NULL);
876 	ASSERT(xfs_iext_valid(ifp, cur));
877 
878 	xfs_iext_inc_seq(ifp);
879 
880 	nr_entries = xfs_iext_leaf_nr_entries(ifp, leaf, cur->pos) - 1;
881 	for (i = cur->pos; i < nr_entries; i++)
882 		leaf->recs[i] = leaf->recs[i + 1];
883 	xfs_iext_rec_clear(&leaf->recs[nr_entries]);
884 	ifp->if_bytes -= sizeof(struct xfs_iext_rec);
885 
886 	if (cur->pos == 0 && nr_entries > 0) {
887 		xfs_iext_update_node(ifp, offset, xfs_iext_leaf_key(leaf, 0), 1,
888 				leaf);
889 		offset = xfs_iext_leaf_key(leaf, 0);
890 	} else if (cur->pos == nr_entries) {
891 		if (ifp->if_height > 1 && leaf->next)
892 			cur->leaf = leaf->next;
893 		else
894 			cur->leaf = NULL;
895 		cur->pos = 0;
896 	}
897 
898 	if (nr_entries >= RECS_PER_LEAF / 2)
899 		return;
900 
901 	if (ifp->if_height > 1)
902 		xfs_iext_rebalance_leaf(ifp, cur, leaf, offset, nr_entries);
903 	else if (nr_entries == 0)
904 		xfs_iext_free_last_leaf(ifp);
905 }
906 
907 /*
908  * Lookup the extent covering bno.
909  *
910  * If there is an extent covering bno return the extent index, and store the
911  * expanded extent structure in *gotp, and the extent cursor in *cur.
912  * If there is no extent covering bno, but there is an extent after it (e.g.
913  * it lies in a hole) return that extent in *gotp and its cursor in *cur
914  * instead.
915  * If bno is beyond the last extent return false, and return an invalid
916  * cursor value.
917  */
918 bool
919 xfs_iext_lookup_extent(
920 	struct xfs_inode	*ip,
921 	struct xfs_ifork	*ifp,
922 	xfs_fileoff_t		offset,
923 	struct xfs_iext_cursor	*cur,
924 	struct xfs_bmbt_irec	*gotp)
925 {
926 	XFS_STATS_INC(ip->i_mount, xs_look_exlist);
927 
928 	cur->leaf = xfs_iext_find_level(ifp, offset, 1);
929 	if (!cur->leaf) {
930 		cur->pos = 0;
931 		return false;
932 	}
933 
934 	for (cur->pos = 0; cur->pos < xfs_iext_max_recs(ifp); cur->pos++) {
935 		struct xfs_iext_rec *rec = cur_rec(cur);
936 
937 		if (xfs_iext_rec_is_empty(rec))
938 			break;
939 		if (xfs_iext_rec_cmp(rec, offset) >= 0)
940 			goto found;
941 	}
942 
943 	/* Try looking in the next node for an entry > offset */
944 	if (ifp->if_height == 1 || !cur->leaf->next)
945 		return false;
946 	cur->leaf = cur->leaf->next;
947 	cur->pos = 0;
948 	if (!xfs_iext_valid(ifp, cur))
949 		return false;
950 found:
951 	xfs_iext_get(gotp, cur_rec(cur));
952 	return true;
953 }
954 
955 /*
956  * Returns the last extent before end, and if this extent doesn't cover
957  * end, update end to the end of the extent.
958  */
959 bool
960 xfs_iext_lookup_extent_before(
961 	struct xfs_inode	*ip,
962 	struct xfs_ifork	*ifp,
963 	xfs_fileoff_t		*end,
964 	struct xfs_iext_cursor	*cur,
965 	struct xfs_bmbt_irec	*gotp)
966 {
967 	/* could be optimized to not even look up the next on a match.. */
968 	if (xfs_iext_lookup_extent(ip, ifp, *end - 1, cur, gotp) &&
969 	    gotp->br_startoff <= *end - 1)
970 		return true;
971 	if (!xfs_iext_prev_extent(ifp, cur, gotp))
972 		return false;
973 	*end = gotp->br_startoff + gotp->br_blockcount;
974 	return true;
975 }
976 
977 void
978 xfs_iext_update_extent(
979 	struct xfs_inode	*ip,
980 	int			state,
981 	struct xfs_iext_cursor	*cur,
982 	struct xfs_bmbt_irec	*new)
983 {
984 	struct xfs_ifork	*ifp = xfs_iext_state_to_fork(ip, state);
985 
986 	xfs_iext_inc_seq(ifp);
987 
988 	if (cur->pos == 0) {
989 		struct xfs_bmbt_irec	old;
990 
991 		xfs_iext_get(&old, cur_rec(cur));
992 		if (new->br_startoff != old.br_startoff) {
993 			xfs_iext_update_node(ifp, old.br_startoff,
994 					new->br_startoff, 1, cur->leaf);
995 		}
996 	}
997 
998 	trace_xfs_bmap_pre_update(ip, cur, state, _RET_IP_);
999 	xfs_iext_set(cur_rec(cur), new);
1000 	trace_xfs_bmap_post_update(ip, cur, state, _RET_IP_);
1001 }
1002 
1003 /*
1004  * Return true if the cursor points at an extent and return the extent structure
1005  * in gotp.  Else return false.
1006  */
1007 bool
1008 xfs_iext_get_extent(
1009 	struct xfs_ifork	*ifp,
1010 	struct xfs_iext_cursor	*cur,
1011 	struct xfs_bmbt_irec	*gotp)
1012 {
1013 	if (!xfs_iext_valid(ifp, cur))
1014 		return false;
1015 	xfs_iext_get(gotp, cur_rec(cur));
1016 	return true;
1017 }
1018 
1019 /*
1020  * This is a recursive function, because of that we need to be extremely
1021  * careful with stack usage.
1022  */
1023 static void
1024 xfs_iext_destroy_node(
1025 	struct xfs_iext_node	*node,
1026 	int			level)
1027 {
1028 	int			i;
1029 
1030 	if (level > 1) {
1031 		for (i = 0; i < KEYS_PER_NODE; i++) {
1032 			if (node->keys[i] == XFS_IEXT_KEY_INVALID)
1033 				break;
1034 			xfs_iext_destroy_node(node->ptrs[i], level - 1);
1035 		}
1036 	}
1037 
1038 	kmem_free(node);
1039 }
1040 
1041 void
1042 xfs_iext_destroy(
1043 	struct xfs_ifork	*ifp)
1044 {
1045 	xfs_iext_destroy_node(ifp->if_u1.if_root, ifp->if_height);
1046 
1047 	ifp->if_bytes = 0;
1048 	ifp->if_height = 0;
1049 	ifp->if_u1.if_root = NULL;
1050 }
1051