2 * Copyright (C) 2011 Red Hat, Inc.
4 * This file is released under the GPL.
7 #include "dm-btree-internal.h"
8 #include "dm-space-map.h"
9 #include "dm-transaction-manager.h"
11 #include <linux/export.h>
12 #include <linux/device-mapper.h>
14 #define DM_MSG_PREFIX "btree"
16 /*----------------------------------------------------------------
18 *--------------------------------------------------------------*/
19 static void memcpy_disk(void *dest
, const void *src
, size_t len
)
20 __dm_written_to_disk(src
)
22 memcpy(dest
, src
, len
);
23 __dm_unbless_for_disk(src
);
26 static void array_insert(void *base
, size_t elt_size
, unsigned nr_elts
,
27 unsigned index
, void *elt
)
28 __dm_written_to_disk(elt
)
31 memmove(base
+ (elt_size
* (index
+ 1)),
32 base
+ (elt_size
* index
),
33 (nr_elts
- index
) * elt_size
);
35 memcpy_disk(base
+ (elt_size
* index
), elt
, elt_size
);
38 /*----------------------------------------------------------------*/
40 /* makes the assumption that no two keys are the same. */
41 static int bsearch(struct node
*n
, uint64_t key
, int want_hi
)
43 int lo
= -1, hi
= le32_to_cpu(n
->header
.nr_entries
);
46 int mid
= lo
+ ((hi
- lo
) / 2);
47 uint64_t mid_key
= le64_to_cpu(n
->keys
[mid
]);
58 return want_hi
? hi
: lo
;
61 int lower_bound(struct node
*n
, uint64_t key
)
63 return bsearch(n
, key
, 0);
66 void inc_children(struct dm_transaction_manager
*tm
, struct node
*n
,
67 struct dm_btree_value_type
*vt
)
70 uint32_t nr_entries
= le32_to_cpu(n
->header
.nr_entries
);
72 if (le32_to_cpu(n
->header
.flags
) & INTERNAL_NODE
)
73 for (i
= 0; i
< nr_entries
; i
++)
74 dm_tm_inc(tm
, value64(n
, i
));
76 for (i
= 0; i
< nr_entries
; i
++)
78 value_ptr(n
, i
, vt
->size
));
81 static int insert_at(size_t value_size
, struct node
*node
, unsigned index
,
82 uint64_t key
, void *value
)
83 __dm_written_to_disk(value
)
85 uint32_t nr_entries
= le32_to_cpu(node
->header
.nr_entries
);
86 __le64 key_le
= cpu_to_le64(key
);
88 if (index
> nr_entries
||
89 index
>= le32_to_cpu(node
->header
.max_entries
)) {
90 DMERR("too many entries in btree node for insert");
91 __dm_unbless_for_disk(value
);
95 __dm_bless_for_disk(&key_le
);
97 array_insert(node
->keys
, sizeof(*node
->keys
), nr_entries
, index
, &key_le
);
98 array_insert(value_base(node
), value_size
, nr_entries
, index
, value
);
99 node
->header
.nr_entries
= cpu_to_le32(nr_entries
+ 1);
104 /*----------------------------------------------------------------*/
107 * We want 3n entries (for some n). This works more nicely for repeated
108 * insert remove loops than (2n + 1).
110 static uint32_t calc_max_entries(size_t value_size
, size_t block_size
)
113 size_t elt_size
= sizeof(uint64_t) + value_size
; /* key + value */
115 block_size
-= sizeof(struct node_header
);
116 total
= block_size
/ elt_size
;
117 n
= total
/ 3; /* rounds down */
122 int dm_btree_empty(struct dm_btree_info
*info
, dm_block_t
*root
)
128 uint32_t max_entries
;
130 r
= new_block(info
, &b
);
134 block_size
= dm_bm_block_size(dm_tm_get_bm(info
->tm
));
135 max_entries
= calc_max_entries(info
->value_type
.size
, block_size
);
137 n
= dm_block_data(b
);
138 memset(n
, 0, block_size
);
139 n
->header
.flags
= cpu_to_le32(LEAF_NODE
);
140 n
->header
.nr_entries
= cpu_to_le32(0);
141 n
->header
.max_entries
= cpu_to_le32(max_entries
);
142 n
->header
.value_size
= cpu_to_le32(info
->value_type
.size
);
144 *root
= dm_block_location(b
);
145 return unlock_block(info
, b
);
147 EXPORT_SYMBOL_GPL(dm_btree_empty
);
149 /*----------------------------------------------------------------*/
152 * Deletion uses a recursive algorithm, since we have limited stack space
153 * we explicitly manage our own stack on the heap.
155 #define MAX_SPINE_DEPTH 64
160 unsigned nr_children
;
161 unsigned current_child
;
165 struct dm_transaction_manager
*tm
;
167 struct frame spine
[MAX_SPINE_DEPTH
];
170 static int top_frame(struct del_stack
*s
, struct frame
**f
)
173 DMERR("btree deletion stack empty");
177 *f
= s
->spine
+ s
->top
;
182 static int unprocessed_frames(struct del_stack
*s
)
187 static int push_frame(struct del_stack
*s
, dm_block_t b
, unsigned level
)
192 if (s
->top
>= MAX_SPINE_DEPTH
- 1) {
193 DMERR("btree deletion stack out of memory");
197 r
= dm_tm_ref(s
->tm
, b
, &ref_count
);
203 * This is a shared node, so we can just decrement it's
204 * reference counter and leave the children.
209 struct frame
*f
= s
->spine
+ ++s
->top
;
211 r
= dm_tm_read_lock(s
->tm
, b
, &btree_node_validator
, &f
->b
);
217 f
->n
= dm_block_data(f
->b
);
219 f
->nr_children
= le32_to_cpu(f
->n
->header
.nr_entries
);
220 f
->current_child
= 0;
226 static void pop_frame(struct del_stack
*s
)
228 struct frame
*f
= s
->spine
+ s
->top
--;
230 dm_tm_dec(s
->tm
, dm_block_location(f
->b
));
231 dm_tm_unlock(s
->tm
, f
->b
);
234 int dm_btree_del(struct dm_btree_info
*info
, dm_block_t root
)
239 s
= kmalloc(sizeof(*s
), GFP_KERNEL
);
245 r
= push_frame(s
, root
, 1);
249 while (unprocessed_frames(s
)) {
254 r
= top_frame(s
, &f
);
258 if (f
->current_child
>= f
->nr_children
) {
263 flags
= le32_to_cpu(f
->n
->header
.flags
);
264 if (flags
& INTERNAL_NODE
) {
265 b
= value64(f
->n
, f
->current_child
);
267 r
= push_frame(s
, b
, f
->level
);
271 } else if (f
->level
!= (info
->levels
- 1)) {
272 b
= value64(f
->n
, f
->current_child
);
274 r
= push_frame(s
, b
, f
->level
+ 1);
279 if (info
->value_type
.dec
) {
282 for (i
= 0; i
< f
->nr_children
; i
++)
283 info
->value_type
.dec(info
->value_type
.context
,
284 value_ptr(f
->n
, i
, info
->value_type
.size
));
286 f
->current_child
= f
->nr_children
;
294 EXPORT_SYMBOL_GPL(dm_btree_del
);
296 /*----------------------------------------------------------------*/
298 static int btree_lookup_raw(struct ro_spine
*s
, dm_block_t block
, uint64_t key
,
299 int (*search_fn
)(struct node
*, uint64_t),
300 uint64_t *result_key
, void *v
, size_t value_size
)
303 uint32_t flags
, nr_entries
;
306 r
= ro_step(s
, block
);
310 i
= search_fn(ro_node(s
), key
);
312 flags
= le32_to_cpu(ro_node(s
)->header
.flags
);
313 nr_entries
= le32_to_cpu(ro_node(s
)->header
.nr_entries
);
314 if (i
< 0 || i
>= nr_entries
)
317 if (flags
& INTERNAL_NODE
)
318 block
= value64(ro_node(s
), i
);
320 } while (!(flags
& LEAF_NODE
));
322 *result_key
= le64_to_cpu(ro_node(s
)->keys
[i
]);
323 memcpy(v
, value_ptr(ro_node(s
), i
, value_size
), value_size
);
328 int dm_btree_lookup(struct dm_btree_info
*info
, dm_block_t root
,
329 uint64_t *keys
, void *value_le
)
331 unsigned level
, last_level
= info
->levels
- 1;
334 __le64 internal_value_le
;
335 struct ro_spine spine
;
337 init_ro_spine(&spine
, info
);
338 for (level
= 0; level
< info
->levels
; level
++) {
342 if (level
== last_level
) {
344 size
= info
->value_type
.size
;
347 value_p
= &internal_value_le
;
348 size
= sizeof(uint64_t);
351 r
= btree_lookup_raw(&spine
, root
, keys
[level
],
356 if (rkey
!= keys
[level
]) {
357 exit_ro_spine(&spine
);
361 exit_ro_spine(&spine
);
365 root
= le64_to_cpu(internal_value_le
);
367 exit_ro_spine(&spine
);
371 EXPORT_SYMBOL_GPL(dm_btree_lookup
);
374 * Splits a node by creating a sibling node and shifting half the nodes
375 * contents across. Assumes there is a parent node, and it has room for
397 * +---------+ +-------+
401 * Where A* is a shadow of A.
403 static int btree_split_sibling(struct shadow_spine
*s
, dm_block_t root
,
404 unsigned parent_index
, uint64_t key
)
408 unsigned nr_left
, nr_right
;
409 struct dm_block
*left
, *right
, *parent
;
410 struct node
*ln
, *rn
, *pn
;
413 left
= shadow_current(s
);
415 r
= new_block(s
->info
, &right
);
419 ln
= dm_block_data(left
);
420 rn
= dm_block_data(right
);
422 nr_left
= le32_to_cpu(ln
->header
.nr_entries
) / 2;
423 nr_right
= le32_to_cpu(ln
->header
.nr_entries
) - nr_left
;
425 ln
->header
.nr_entries
= cpu_to_le32(nr_left
);
427 rn
->header
.flags
= ln
->header
.flags
;
428 rn
->header
.nr_entries
= cpu_to_le32(nr_right
);
429 rn
->header
.max_entries
= ln
->header
.max_entries
;
430 rn
->header
.value_size
= ln
->header
.value_size
;
431 memcpy(rn
->keys
, ln
->keys
+ nr_left
, nr_right
* sizeof(rn
->keys
[0]));
433 size
= le32_to_cpu(ln
->header
.flags
) & INTERNAL_NODE
?
434 sizeof(uint64_t) : s
->info
->value_type
.size
;
435 memcpy(value_ptr(rn
, 0, size
), value_ptr(ln
, nr_left
, size
),
439 * Patch up the parent
441 parent
= shadow_parent(s
);
443 pn
= dm_block_data(parent
);
444 location
= cpu_to_le64(dm_block_location(left
));
445 __dm_bless_for_disk(&location
);
446 memcpy_disk(value_ptr(pn
, parent_index
, sizeof(__le64
)),
447 &location
, sizeof(__le64
));
449 location
= cpu_to_le64(dm_block_location(right
));
450 __dm_bless_for_disk(&location
);
452 r
= insert_at(sizeof(__le64
), pn
, parent_index
+ 1,
453 le64_to_cpu(rn
->keys
[0]), &location
);
457 if (key
< le64_to_cpu(rn
->keys
[0])) {
458 unlock_block(s
->info
, right
);
461 unlock_block(s
->info
, left
);
469 * Splits a node by creating two new children beneath the given node.
485 * +-------+ +-------+
486 * | B +++ | | C +++ |
487 * +-------+ +-------+
489 static int btree_split_beneath(struct shadow_spine
*s
, uint64_t key
)
493 unsigned nr_left
, nr_right
;
494 struct dm_block
*left
, *right
, *new_parent
;
495 struct node
*pn
, *ln
, *rn
;
498 new_parent
= shadow_current(s
);
500 r
= new_block(s
->info
, &left
);
504 r
= new_block(s
->info
, &right
);
506 /* FIXME: put left */
510 pn
= dm_block_data(new_parent
);
511 ln
= dm_block_data(left
);
512 rn
= dm_block_data(right
);
514 nr_left
= le32_to_cpu(pn
->header
.nr_entries
) / 2;
515 nr_right
= le32_to_cpu(pn
->header
.nr_entries
) - nr_left
;
517 ln
->header
.flags
= pn
->header
.flags
;
518 ln
->header
.nr_entries
= cpu_to_le32(nr_left
);
519 ln
->header
.max_entries
= pn
->header
.max_entries
;
520 ln
->header
.value_size
= pn
->header
.value_size
;
522 rn
->header
.flags
= pn
->header
.flags
;
523 rn
->header
.nr_entries
= cpu_to_le32(nr_right
);
524 rn
->header
.max_entries
= pn
->header
.max_entries
;
525 rn
->header
.value_size
= pn
->header
.value_size
;
527 memcpy(ln
->keys
, pn
->keys
, nr_left
* sizeof(pn
->keys
[0]));
528 memcpy(rn
->keys
, pn
->keys
+ nr_left
, nr_right
* sizeof(pn
->keys
[0]));
530 size
= le32_to_cpu(pn
->header
.flags
) & INTERNAL_NODE
?
531 sizeof(__le64
) : s
->info
->value_type
.size
;
532 memcpy(value_ptr(ln
, 0, size
), value_ptr(pn
, 0, size
), nr_left
* size
);
533 memcpy(value_ptr(rn
, 0, size
), value_ptr(pn
, nr_left
, size
),
536 /* new_parent should just point to l and r now */
537 pn
->header
.flags
= cpu_to_le32(INTERNAL_NODE
);
538 pn
->header
.nr_entries
= cpu_to_le32(2);
539 pn
->header
.max_entries
= cpu_to_le32(
540 calc_max_entries(sizeof(__le64
),
542 dm_tm_get_bm(s
->info
->tm
))));
543 pn
->header
.value_size
= cpu_to_le32(sizeof(__le64
));
545 val
= cpu_to_le64(dm_block_location(left
));
546 __dm_bless_for_disk(&val
);
547 pn
->keys
[0] = ln
->keys
[0];
548 memcpy_disk(value_ptr(pn
, 0, sizeof(__le64
)), &val
, sizeof(__le64
));
550 val
= cpu_to_le64(dm_block_location(right
));
551 __dm_bless_for_disk(&val
);
552 pn
->keys
[1] = rn
->keys
[0];
553 memcpy_disk(value_ptr(pn
, 1, sizeof(__le64
)), &val
, sizeof(__le64
));
556 * rejig the spine. This is ugly, since it knows too
557 * much about the spine
559 if (s
->nodes
[0] != new_parent
) {
560 unlock_block(s
->info
, s
->nodes
[0]);
561 s
->nodes
[0] = new_parent
;
563 if (key
< le64_to_cpu(rn
->keys
[0])) {
564 unlock_block(s
->info
, right
);
567 unlock_block(s
->info
, left
);
575 static int btree_insert_raw(struct shadow_spine
*s
, dm_block_t root
,
576 struct dm_btree_value_type
*vt
,
577 uint64_t key
, unsigned *index
)
579 int r
, i
= *index
, top
= 1;
583 r
= shadow_step(s
, root
, vt
);
587 node
= dm_block_data(shadow_current(s
));
590 * We have to patch up the parent node, ugly, but I don't
591 * see a way to do this automatically as part of the spine
594 if (shadow_has_parent(s
) && i
>= 0) { /* FIXME: second clause unness. */
595 __le64 location
= cpu_to_le64(dm_block_location(shadow_current(s
)));
597 __dm_bless_for_disk(&location
);
598 memcpy_disk(value_ptr(dm_block_data(shadow_parent(s
)), i
, sizeof(uint64_t)),
599 &location
, sizeof(__le64
));
602 node
= dm_block_data(shadow_current(s
));
604 if (node
->header
.nr_entries
== node
->header
.max_entries
) {
606 r
= btree_split_beneath(s
, key
);
608 r
= btree_split_sibling(s
, root
, i
, key
);
614 node
= dm_block_data(shadow_current(s
));
616 i
= lower_bound(node
, key
);
618 if (le32_to_cpu(node
->header
.flags
) & LEAF_NODE
)
622 /* change the bounds on the lowest key */
623 node
->keys
[0] = cpu_to_le64(key
);
627 root
= value64(node
, i
);
631 if (i
< 0 || le64_to_cpu(node
->keys
[i
]) != key
)
638 static int insert(struct dm_btree_info
*info
, dm_block_t root
,
639 uint64_t *keys
, void *value
, dm_block_t
*new_root
,
641 __dm_written_to_disk(value
)
644 unsigned level
, index
= -1, last_level
= info
->levels
- 1;
645 dm_block_t block
= root
;
646 struct shadow_spine spine
;
648 struct dm_btree_value_type le64_type
;
650 le64_type
.context
= NULL
;
651 le64_type
.size
= sizeof(__le64
);
652 le64_type
.inc
= NULL
;
653 le64_type
.dec
= NULL
;
654 le64_type
.equal
= NULL
;
656 init_shadow_spine(&spine
, info
);
658 for (level
= 0; level
< (info
->levels
- 1); level
++) {
659 r
= btree_insert_raw(&spine
, block
, &le64_type
, keys
[level
], &index
);
663 n
= dm_block_data(shadow_current(&spine
));
664 need_insert
= ((index
>= le32_to_cpu(n
->header
.nr_entries
)) ||
665 (le64_to_cpu(n
->keys
[index
]) != keys
[level
]));
671 r
= dm_btree_empty(info
, &new_tree
);
675 new_le
= cpu_to_le64(new_tree
);
676 __dm_bless_for_disk(&new_le
);
678 r
= insert_at(sizeof(uint64_t), n
, index
,
679 keys
[level
], &new_le
);
684 if (level
< last_level
)
685 block
= value64(n
, index
);
688 r
= btree_insert_raw(&spine
, block
, &info
->value_type
,
689 keys
[level
], &index
);
693 n
= dm_block_data(shadow_current(&spine
));
694 need_insert
= ((index
>= le32_to_cpu(n
->header
.nr_entries
)) ||
695 (le64_to_cpu(n
->keys
[index
]) != keys
[level
]));
701 r
= insert_at(info
->value_type
.size
, n
, index
,
709 if (info
->value_type
.dec
&&
710 (!info
->value_type
.equal
||
711 !info
->value_type
.equal(
712 info
->value_type
.context
,
713 value_ptr(n
, index
, info
->value_type
.size
),
715 info
->value_type
.dec(info
->value_type
.context
,
716 value_ptr(n
, index
, info
->value_type
.size
));
718 memcpy_disk(value_ptr(n
, index
, info
->value_type
.size
),
719 value
, info
->value_type
.size
);
722 *new_root
= shadow_root(&spine
);
723 exit_shadow_spine(&spine
);
728 __dm_unbless_for_disk(value
);
730 exit_shadow_spine(&spine
);
734 int dm_btree_insert(struct dm_btree_info
*info
, dm_block_t root
,
735 uint64_t *keys
, void *value
, dm_block_t
*new_root
)
736 __dm_written_to_disk(value
)
738 return insert(info
, root
, keys
, value
, new_root
, NULL
);
740 EXPORT_SYMBOL_GPL(dm_btree_insert
);
742 int dm_btree_insert_notify(struct dm_btree_info
*info
, dm_block_t root
,
743 uint64_t *keys
, void *value
, dm_block_t
*new_root
,
745 __dm_written_to_disk(value
)
747 return insert(info
, root
, keys
, value
, new_root
, inserted
);
749 EXPORT_SYMBOL_GPL(dm_btree_insert_notify
);
751 /*----------------------------------------------------------------*/
753 static int find_highest_key(struct ro_spine
*s
, dm_block_t block
,
754 uint64_t *result_key
, dm_block_t
*next_block
)
760 r
= ro_step(s
, block
);
764 flags
= le32_to_cpu(ro_node(s
)->header
.flags
);
765 i
= le32_to_cpu(ro_node(s
)->header
.nr_entries
);
771 *result_key
= le64_to_cpu(ro_node(s
)->keys
[i
]);
772 if (next_block
|| flags
& INTERNAL_NODE
)
773 block
= value64(ro_node(s
), i
);
775 } while (flags
& INTERNAL_NODE
);
782 int dm_btree_find_highest_key(struct dm_btree_info
*info
, dm_block_t root
,
783 uint64_t *result_keys
)
785 int r
= 0, count
= 0, level
;
786 struct ro_spine spine
;
788 init_ro_spine(&spine
, info
);
789 for (level
= 0; level
< info
->levels
; level
++) {
790 r
= find_highest_key(&spine
, root
, result_keys
+ level
,
791 level
== info
->levels
- 1 ? NULL
: &root
);
801 exit_ro_spine(&spine
);
803 return r
? r
: count
;
805 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key
);