1 /*
2 * Copyright (C) 2011 Red Hat, Inc.
3 *
4 * This file is released under the GPL.
5 */
6
7 #include "dm-btree-internal.h"
8 #include "dm-space-map.h"
9 #include "dm-transaction-manager.h"
10
11 #include <linux/export.h>
12 #include <linux/device-mapper.h>
13
14 #define DM_MSG_PREFIX "btree"
15
16 /*----------------------------------------------------------------
17 * Array manipulation
18 *--------------------------------------------------------------*/
memcpy_disk(void * dest,const void * src,size_t len)19 static void memcpy_disk(void *dest, const void *src, size_t len)
20 __dm_written_to_disk(src)
21 {
22 memcpy(dest, src, len);
23 __dm_unbless_for_disk(src);
24 }
25
array_insert(void * base,size_t elt_size,unsigned nr_elts,unsigned index,void * elt)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)
29 {
30 if (index < nr_elts)
31 memmove(base + (elt_size * (index + 1)),
32 base + (elt_size * index),
33 (nr_elts - index) * elt_size);
34
35 memcpy_disk(base + (elt_size * index), elt, elt_size);
36 }
37
38 /*----------------------------------------------------------------*/
39
40 /* makes the assumption that no two keys are the same. */
bsearch(struct btree_node * n,uint64_t key,int want_hi)41 static int bsearch(struct btree_node *n, uint64_t key, int want_hi)
42 {
43 int lo = -1, hi = le32_to_cpu(n->header.nr_entries);
44
45 while (hi - lo > 1) {
46 int mid = lo + ((hi - lo) / 2);
47 uint64_t mid_key = le64_to_cpu(n->keys[mid]);
48
49 if (mid_key == key)
50 return mid;
51
52 if (mid_key < key)
53 lo = mid;
54 else
55 hi = mid;
56 }
57
58 return want_hi ? hi : lo;
59 }
60
lower_bound(struct btree_node * n,uint64_t key)61 int lower_bound(struct btree_node *n, uint64_t key)
62 {
63 return bsearch(n, key, 0);
64 }
65
upper_bound(struct btree_node * n,uint64_t key)66 static int upper_bound(struct btree_node *n, uint64_t key)
67 {
68 return bsearch(n, key, 1);
69 }
70
inc_children(struct dm_transaction_manager * tm,struct btree_node * n,struct dm_btree_value_type * vt)71 void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
72 struct dm_btree_value_type *vt)
73 {
74 unsigned i;
75 uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
76
77 if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
78 for (i = 0; i < nr_entries; i++)
79 dm_tm_inc(tm, value64(n, i));
80 else if (vt->inc)
81 for (i = 0; i < nr_entries; i++)
82 vt->inc(vt->context, value_ptr(n, i));
83 }
84
insert_at(size_t value_size,struct btree_node * node,unsigned index,uint64_t key,void * value)85 static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
86 uint64_t key, void *value)
87 __dm_written_to_disk(value)
88 {
89 uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
90 __le64 key_le = cpu_to_le64(key);
91
92 if (index > nr_entries ||
93 index >= le32_to_cpu(node->header.max_entries)) {
94 DMERR("too many entries in btree node for insert");
95 __dm_unbless_for_disk(value);
96 return -ENOMEM;
97 }
98
99 __dm_bless_for_disk(&key_le);
100
101 array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
102 array_insert(value_base(node), value_size, nr_entries, index, value);
103 node->header.nr_entries = cpu_to_le32(nr_entries + 1);
104
105 return 0;
106 }
107
108 /*----------------------------------------------------------------*/
109
110 /*
111 * We want 3n entries (for some n). This works more nicely for repeated
112 * insert remove loops than (2n + 1).
113 */
calc_max_entries(size_t value_size,size_t block_size)114 static uint32_t calc_max_entries(size_t value_size, size_t block_size)
115 {
116 uint32_t total, n;
117 size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
118
119 block_size -= sizeof(struct node_header);
120 total = block_size / elt_size;
121 n = total / 3; /* rounds down */
122
123 return 3 * n;
124 }
125
dm_btree_empty(struct dm_btree_info * info,dm_block_t * root)126 int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
127 {
128 int r;
129 struct dm_block *b;
130 struct btree_node *n;
131 size_t block_size;
132 uint32_t max_entries;
133
134 r = new_block(info, &b);
135 if (r < 0)
136 return r;
137
138 block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
139 max_entries = calc_max_entries(info->value_type.size, block_size);
140
141 n = dm_block_data(b);
142 memset(n, 0, block_size);
143 n->header.flags = cpu_to_le32(LEAF_NODE);
144 n->header.nr_entries = cpu_to_le32(0);
145 n->header.max_entries = cpu_to_le32(max_entries);
146 n->header.value_size = cpu_to_le32(info->value_type.size);
147
148 *root = dm_block_location(b);
149 unlock_block(info, b);
150
151 return 0;
152 }
153 EXPORT_SYMBOL_GPL(dm_btree_empty);
154
155 /*----------------------------------------------------------------*/
156
157 /*
158 * Deletion uses a recursive algorithm, since we have limited stack space
159 * we explicitly manage our own stack on the heap.
160 */
161 #define MAX_SPINE_DEPTH 64
162 struct frame {
163 struct dm_block *b;
164 struct btree_node *n;
165 unsigned level;
166 unsigned nr_children;
167 unsigned current_child;
168 };
169
170 struct del_stack {
171 struct dm_btree_info *info;
172 struct dm_transaction_manager *tm;
173 int top;
174 struct frame spine[MAX_SPINE_DEPTH];
175 };
176
top_frame(struct del_stack * s,struct frame ** f)177 static int top_frame(struct del_stack *s, struct frame **f)
178 {
179 if (s->top < 0) {
180 DMERR("btree deletion stack empty");
181 return -EINVAL;
182 }
183
184 *f = s->spine + s->top;
185
186 return 0;
187 }
188
unprocessed_frames(struct del_stack * s)189 static int unprocessed_frames(struct del_stack *s)
190 {
191 return s->top >= 0;
192 }
193
prefetch_children(struct del_stack * s,struct frame * f)194 static void prefetch_children(struct del_stack *s, struct frame *f)
195 {
196 unsigned i;
197 struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
198
199 for (i = 0; i < f->nr_children; i++)
200 dm_bm_prefetch(bm, value64(f->n, i));
201 }
202
is_internal_level(struct dm_btree_info * info,struct frame * f)203 static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
204 {
205 return f->level < (info->levels - 1);
206 }
207
push_frame(struct del_stack * s,dm_block_t b,unsigned level)208 static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
209 {
210 int r;
211 uint32_t ref_count;
212
213 if (s->top >= MAX_SPINE_DEPTH - 1) {
214 DMERR("btree deletion stack out of memory");
215 return -ENOMEM;
216 }
217
218 r = dm_tm_ref(s->tm, b, &ref_count);
219 if (r)
220 return r;
221
222 if (ref_count > 1)
223 /*
224 * This is a shared node, so we can just decrement it's
225 * reference counter and leave the children.
226 */
227 dm_tm_dec(s->tm, b);
228
229 else {
230 uint32_t flags;
231 struct frame *f = s->spine + ++s->top;
232
233 r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
234 if (r) {
235 s->top--;
236 return r;
237 }
238
239 f->n = dm_block_data(f->b);
240 f->level = level;
241 f->nr_children = le32_to_cpu(f->n->header.nr_entries);
242 f->current_child = 0;
243
244 flags = le32_to_cpu(f->n->header.flags);
245 if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
246 prefetch_children(s, f);
247 }
248
249 return 0;
250 }
251
pop_frame(struct del_stack * s)252 static void pop_frame(struct del_stack *s)
253 {
254 struct frame *f = s->spine + s->top--;
255
256 dm_tm_dec(s->tm, dm_block_location(f->b));
257 dm_tm_unlock(s->tm, f->b);
258 }
259
unlock_all_frames(struct del_stack * s)260 static void unlock_all_frames(struct del_stack *s)
261 {
262 struct frame *f;
263
264 while (unprocessed_frames(s)) {
265 f = s->spine + s->top--;
266 dm_tm_unlock(s->tm, f->b);
267 }
268 }
269
dm_btree_del(struct dm_btree_info * info,dm_block_t root)270 int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
271 {
272 int r;
273 struct del_stack *s;
274
275 s = kmalloc(sizeof(*s), GFP_NOIO);
276 if (!s)
277 return -ENOMEM;
278 s->info = info;
279 s->tm = info->tm;
280 s->top = -1;
281
282 r = push_frame(s, root, 0);
283 if (r)
284 goto out;
285
286 while (unprocessed_frames(s)) {
287 uint32_t flags;
288 struct frame *f;
289 dm_block_t b;
290
291 r = top_frame(s, &f);
292 if (r)
293 goto out;
294
295 if (f->current_child >= f->nr_children) {
296 pop_frame(s);
297 continue;
298 }
299
300 flags = le32_to_cpu(f->n->header.flags);
301 if (flags & INTERNAL_NODE) {
302 b = value64(f->n, f->current_child);
303 f->current_child++;
304 r = push_frame(s, b, f->level);
305 if (r)
306 goto out;
307
308 } else if (is_internal_level(info, f)) {
309 b = value64(f->n, f->current_child);
310 f->current_child++;
311 r = push_frame(s, b, f->level + 1);
312 if (r)
313 goto out;
314
315 } else {
316 if (info->value_type.dec) {
317 unsigned i;
318
319 for (i = 0; i < f->nr_children; i++)
320 info->value_type.dec(info->value_type.context,
321 value_ptr(f->n, i));
322 }
323 pop_frame(s);
324 }
325 }
326 out:
327 if (r) {
328 /* cleanup all frames of del_stack */
329 unlock_all_frames(s);
330 }
331 kfree(s);
332
333 return r;
334 }
335 EXPORT_SYMBOL_GPL(dm_btree_del);
336
337 /*----------------------------------------------------------------*/
338
btree_lookup_raw(struct ro_spine * s,dm_block_t block,uint64_t key,int (* search_fn)(struct btree_node *,uint64_t),uint64_t * result_key,void * v,size_t value_size)339 static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
340 int (*search_fn)(struct btree_node *, uint64_t),
341 uint64_t *result_key, void *v, size_t value_size)
342 {
343 int i, r;
344 uint32_t flags, nr_entries;
345
346 do {
347 r = ro_step(s, block);
348 if (r < 0)
349 return r;
350
351 i = search_fn(ro_node(s), key);
352
353 flags = le32_to_cpu(ro_node(s)->header.flags);
354 nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
355 if (i < 0 || i >= nr_entries)
356 return -ENODATA;
357
358 if (flags & INTERNAL_NODE)
359 block = value64(ro_node(s), i);
360
361 } while (!(flags & LEAF_NODE));
362
363 *result_key = le64_to_cpu(ro_node(s)->keys[i]);
364 memcpy(v, value_ptr(ro_node(s), i), value_size);
365
366 return 0;
367 }
368
dm_btree_lookup(struct dm_btree_info * info,dm_block_t root,uint64_t * keys,void * value_le)369 int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
370 uint64_t *keys, void *value_le)
371 {
372 unsigned level, last_level = info->levels - 1;
373 int r = -ENODATA;
374 uint64_t rkey;
375 __le64 internal_value_le;
376 struct ro_spine spine;
377
378 init_ro_spine(&spine, info);
379 for (level = 0; level < info->levels; level++) {
380 size_t size;
381 void *value_p;
382
383 if (level == last_level) {
384 value_p = value_le;
385 size = info->value_type.size;
386
387 } else {
388 value_p = &internal_value_le;
389 size = sizeof(uint64_t);
390 }
391
392 r = btree_lookup_raw(&spine, root, keys[level],
393 lower_bound, &rkey,
394 value_p, size);
395
396 if (!r) {
397 if (rkey != keys[level]) {
398 exit_ro_spine(&spine);
399 return -ENODATA;
400 }
401 } else {
402 exit_ro_spine(&spine);
403 return r;
404 }
405
406 root = le64_to_cpu(internal_value_le);
407 }
408 exit_ro_spine(&spine);
409
410 return r;
411 }
412 EXPORT_SYMBOL_GPL(dm_btree_lookup);
413
dm_btree_lookup_next_single(struct dm_btree_info * info,dm_block_t root,uint64_t key,uint64_t * rkey,void * value_le)414 static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
415 uint64_t key, uint64_t *rkey, void *value_le)
416 {
417 int r, i;
418 uint32_t flags, nr_entries;
419 struct dm_block *node;
420 struct btree_node *n;
421
422 r = bn_read_lock(info, root, &node);
423 if (r)
424 return r;
425
426 n = dm_block_data(node);
427 flags = le32_to_cpu(n->header.flags);
428 nr_entries = le32_to_cpu(n->header.nr_entries);
429
430 if (flags & INTERNAL_NODE) {
431 i = lower_bound(n, key);
432 if (i < 0) {
433 /*
434 * avoid early -ENODATA return when all entries are
435 * higher than the search @key.
436 */
437 i = 0;
438 }
439 if (i >= nr_entries) {
440 r = -ENODATA;
441 goto out;
442 }
443
444 r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
445 if (r == -ENODATA && i < (nr_entries - 1)) {
446 i++;
447 r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
448 }
449
450 } else {
451 i = upper_bound(n, key);
452 if (i < 0 || i >= nr_entries) {
453 r = -ENODATA;
454 goto out;
455 }
456
457 *rkey = le64_to_cpu(n->keys[i]);
458 memcpy(value_le, value_ptr(n, i), info->value_type.size);
459 }
460 out:
461 dm_tm_unlock(info->tm, node);
462 return r;
463 }
464
dm_btree_lookup_next(struct dm_btree_info * info,dm_block_t root,uint64_t * keys,uint64_t * rkey,void * value_le)465 int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
466 uint64_t *keys, uint64_t *rkey, void *value_le)
467 {
468 unsigned level;
469 int r = -ENODATA;
470 __le64 internal_value_le;
471 struct ro_spine spine;
472
473 init_ro_spine(&spine, info);
474 for (level = 0; level < info->levels - 1u; level++) {
475 r = btree_lookup_raw(&spine, root, keys[level],
476 lower_bound, rkey,
477 &internal_value_le, sizeof(uint64_t));
478 if (r)
479 goto out;
480
481 if (*rkey != keys[level]) {
482 r = -ENODATA;
483 goto out;
484 }
485
486 root = le64_to_cpu(internal_value_le);
487 }
488
489 r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
490 out:
491 exit_ro_spine(&spine);
492 return r;
493 }
494
495 EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
496
497 /*
498 * Splits a node by creating a sibling node and shifting half the nodes
499 * contents across. Assumes there is a parent node, and it has room for
500 * another child.
501 *
502 * Before:
503 * +--------+
504 * | Parent |
505 * +--------+
506 * |
507 * v
508 * +----------+
509 * | A ++++++ |
510 * +----------+
511 *
512 *
513 * After:
514 * +--------+
515 * | Parent |
516 * +--------+
517 * | |
518 * v +------+
519 * +---------+ |
520 * | A* +++ | v
521 * +---------+ +-------+
522 * | B +++ |
523 * +-------+
524 *
525 * Where A* is a shadow of A.
526 */
btree_split_sibling(struct shadow_spine * s,unsigned parent_index,uint64_t key)527 static int btree_split_sibling(struct shadow_spine *s, unsigned parent_index,
528 uint64_t key)
529 {
530 int r;
531 size_t size;
532 unsigned nr_left, nr_right;
533 struct dm_block *left, *right, *parent;
534 struct btree_node *ln, *rn, *pn;
535 __le64 location;
536
537 left = shadow_current(s);
538
539 r = new_block(s->info, &right);
540 if (r < 0)
541 return r;
542
543 ln = dm_block_data(left);
544 rn = dm_block_data(right);
545
546 nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
547 nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
548
549 ln->header.nr_entries = cpu_to_le32(nr_left);
550
551 rn->header.flags = ln->header.flags;
552 rn->header.nr_entries = cpu_to_le32(nr_right);
553 rn->header.max_entries = ln->header.max_entries;
554 rn->header.value_size = ln->header.value_size;
555 memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
556
557 size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
558 sizeof(uint64_t) : s->info->value_type.size;
559 memcpy(value_ptr(rn, 0), value_ptr(ln, nr_left),
560 size * nr_right);
561
562 /*
563 * Patch up the parent
564 */
565 parent = shadow_parent(s);
566
567 pn = dm_block_data(parent);
568 location = cpu_to_le64(dm_block_location(left));
569 __dm_bless_for_disk(&location);
570 memcpy_disk(value_ptr(pn, parent_index),
571 &location, sizeof(__le64));
572
573 location = cpu_to_le64(dm_block_location(right));
574 __dm_bless_for_disk(&location);
575
576 r = insert_at(sizeof(__le64), pn, parent_index + 1,
577 le64_to_cpu(rn->keys[0]), &location);
578 if (r) {
579 unlock_block(s->info, right);
580 return r;
581 }
582
583 if (key < le64_to_cpu(rn->keys[0])) {
584 unlock_block(s->info, right);
585 s->nodes[1] = left;
586 } else {
587 unlock_block(s->info, left);
588 s->nodes[1] = right;
589 }
590
591 return 0;
592 }
593
594 /*
595 * Splits a node by creating two new children beneath the given node.
596 *
597 * Before:
598 * +----------+
599 * | A ++++++ |
600 * +----------+
601 *
602 *
603 * After:
604 * +------------+
605 * | A (shadow) |
606 * +------------+
607 * | |
608 * +------+ +----+
609 * | |
610 * v v
611 * +-------+ +-------+
612 * | B +++ | | C +++ |
613 * +-------+ +-------+
614 */
btree_split_beneath(struct shadow_spine * s,uint64_t key)615 static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
616 {
617 int r;
618 size_t size;
619 unsigned nr_left, nr_right;
620 struct dm_block *left, *right, *new_parent;
621 struct btree_node *pn, *ln, *rn;
622 __le64 val;
623
624 new_parent = shadow_current(s);
625
626 r = new_block(s->info, &left);
627 if (r < 0)
628 return r;
629
630 r = new_block(s->info, &right);
631 if (r < 0) {
632 unlock_block(s->info, left);
633 return r;
634 }
635
636 pn = dm_block_data(new_parent);
637 ln = dm_block_data(left);
638 rn = dm_block_data(right);
639
640 nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
641 nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
642
643 ln->header.flags = pn->header.flags;
644 ln->header.nr_entries = cpu_to_le32(nr_left);
645 ln->header.max_entries = pn->header.max_entries;
646 ln->header.value_size = pn->header.value_size;
647
648 rn->header.flags = pn->header.flags;
649 rn->header.nr_entries = cpu_to_le32(nr_right);
650 rn->header.max_entries = pn->header.max_entries;
651 rn->header.value_size = pn->header.value_size;
652
653 memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
654 memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
655
656 size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
657 sizeof(__le64) : s->info->value_type.size;
658 memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
659 memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
660 nr_right * size);
661
662 /* new_parent should just point to l and r now */
663 pn->header.flags = cpu_to_le32(INTERNAL_NODE);
664 pn->header.nr_entries = cpu_to_le32(2);
665 pn->header.max_entries = cpu_to_le32(
666 calc_max_entries(sizeof(__le64),
667 dm_bm_block_size(
668 dm_tm_get_bm(s->info->tm))));
669 pn->header.value_size = cpu_to_le32(sizeof(__le64));
670
671 val = cpu_to_le64(dm_block_location(left));
672 __dm_bless_for_disk(&val);
673 pn->keys[0] = ln->keys[0];
674 memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
675
676 val = cpu_to_le64(dm_block_location(right));
677 __dm_bless_for_disk(&val);
678 pn->keys[1] = rn->keys[0];
679 memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
680
681 unlock_block(s->info, left);
682 unlock_block(s->info, right);
683 return 0;
684 }
685
btree_insert_raw(struct shadow_spine * s,dm_block_t root,struct dm_btree_value_type * vt,uint64_t key,unsigned * index)686 static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
687 struct dm_btree_value_type *vt,
688 uint64_t key, unsigned *index)
689 {
690 int r, i = *index, top = 1;
691 struct btree_node *node;
692
693 for (;;) {
694 r = shadow_step(s, root, vt);
695 if (r < 0)
696 return r;
697
698 node = dm_block_data(shadow_current(s));
699
700 /*
701 * We have to patch up the parent node, ugly, but I don't
702 * see a way to do this automatically as part of the spine
703 * op.
704 */
705 if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
706 __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
707
708 __dm_bless_for_disk(&location);
709 memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
710 &location, sizeof(__le64));
711 }
712
713 node = dm_block_data(shadow_current(s));
714
715 if (node->header.nr_entries == node->header.max_entries) {
716 if (top)
717 r = btree_split_beneath(s, key);
718 else
719 r = btree_split_sibling(s, i, key);
720
721 if (r < 0)
722 return r;
723 }
724
725 node = dm_block_data(shadow_current(s));
726
727 i = lower_bound(node, key);
728
729 if (le32_to_cpu(node->header.flags) & LEAF_NODE)
730 break;
731
732 if (i < 0) {
733 /* change the bounds on the lowest key */
734 node->keys[0] = cpu_to_le64(key);
735 i = 0;
736 }
737
738 root = value64(node, i);
739 top = 0;
740 }
741
742 if (i < 0 || le64_to_cpu(node->keys[i]) != key)
743 i++;
744
745 *index = i;
746 return 0;
747 }
748
need_insert(struct btree_node * node,uint64_t * keys,unsigned level,unsigned index)749 static bool need_insert(struct btree_node *node, uint64_t *keys,
750 unsigned level, unsigned index)
751 {
752 return ((index >= le32_to_cpu(node->header.nr_entries)) ||
753 (le64_to_cpu(node->keys[index]) != keys[level]));
754 }
755
insert(struct dm_btree_info * info,dm_block_t root,uint64_t * keys,void * value,dm_block_t * new_root,int * inserted)756 static int insert(struct dm_btree_info *info, dm_block_t root,
757 uint64_t *keys, void *value, dm_block_t *new_root,
758 int *inserted)
759 __dm_written_to_disk(value)
760 {
761 int r;
762 unsigned level, index = -1, last_level = info->levels - 1;
763 dm_block_t block = root;
764 struct shadow_spine spine;
765 struct btree_node *n;
766 struct dm_btree_value_type le64_type;
767
768 init_le64_type(info->tm, &le64_type);
769 init_shadow_spine(&spine, info);
770
771 for (level = 0; level < (info->levels - 1); level++) {
772 r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
773 if (r < 0)
774 goto bad;
775
776 n = dm_block_data(shadow_current(&spine));
777
778 if (need_insert(n, keys, level, index)) {
779 dm_block_t new_tree;
780 __le64 new_le;
781
782 r = dm_btree_empty(info, &new_tree);
783 if (r < 0)
784 goto bad;
785
786 new_le = cpu_to_le64(new_tree);
787 __dm_bless_for_disk(&new_le);
788
789 r = insert_at(sizeof(uint64_t), n, index,
790 keys[level], &new_le);
791 if (r)
792 goto bad;
793 }
794
795 if (level < last_level)
796 block = value64(n, index);
797 }
798
799 r = btree_insert_raw(&spine, block, &info->value_type,
800 keys[level], &index);
801 if (r < 0)
802 goto bad;
803
804 n = dm_block_data(shadow_current(&spine));
805
806 if (need_insert(n, keys, level, index)) {
807 if (inserted)
808 *inserted = 1;
809
810 r = insert_at(info->value_type.size, n, index,
811 keys[level], value);
812 if (r)
813 goto bad_unblessed;
814 } else {
815 if (inserted)
816 *inserted = 0;
817
818 if (info->value_type.dec &&
819 (!info->value_type.equal ||
820 !info->value_type.equal(
821 info->value_type.context,
822 value_ptr(n, index),
823 value))) {
824 info->value_type.dec(info->value_type.context,
825 value_ptr(n, index));
826 }
827 memcpy_disk(value_ptr(n, index),
828 value, info->value_type.size);
829 }
830
831 *new_root = shadow_root(&spine);
832 exit_shadow_spine(&spine);
833
834 return 0;
835
836 bad:
837 __dm_unbless_for_disk(value);
838 bad_unblessed:
839 exit_shadow_spine(&spine);
840 return r;
841 }
842
dm_btree_insert(struct dm_btree_info * info,dm_block_t root,uint64_t * keys,void * value,dm_block_t * new_root)843 int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
844 uint64_t *keys, void *value, dm_block_t *new_root)
845 __dm_written_to_disk(value)
846 {
847 return insert(info, root, keys, value, new_root, NULL);
848 }
849 EXPORT_SYMBOL_GPL(dm_btree_insert);
850
dm_btree_insert_notify(struct dm_btree_info * info,dm_block_t root,uint64_t * keys,void * value,dm_block_t * new_root,int * inserted)851 int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
852 uint64_t *keys, void *value, dm_block_t *new_root,
853 int *inserted)
854 __dm_written_to_disk(value)
855 {
856 return insert(info, root, keys, value, new_root, inserted);
857 }
858 EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
859
860 /*----------------------------------------------------------------*/
861
find_key(struct ro_spine * s,dm_block_t block,bool find_highest,uint64_t * result_key,dm_block_t * next_block)862 static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
863 uint64_t *result_key, dm_block_t *next_block)
864 {
865 int i, r;
866 uint32_t flags;
867
868 do {
869 r = ro_step(s, block);
870 if (r < 0)
871 return r;
872
873 flags = le32_to_cpu(ro_node(s)->header.flags);
874 i = le32_to_cpu(ro_node(s)->header.nr_entries);
875 if (!i)
876 return -ENODATA;
877 else
878 i--;
879
880 if (find_highest)
881 *result_key = le64_to_cpu(ro_node(s)->keys[i]);
882 else
883 *result_key = le64_to_cpu(ro_node(s)->keys[0]);
884
885 if (next_block || flags & INTERNAL_NODE) {
886 if (find_highest)
887 block = value64(ro_node(s), i);
888 else
889 block = value64(ro_node(s), 0);
890 }
891
892 } while (flags & INTERNAL_NODE);
893
894 if (next_block)
895 *next_block = block;
896 return 0;
897 }
898
dm_btree_find_key(struct dm_btree_info * info,dm_block_t root,bool find_highest,uint64_t * result_keys)899 static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
900 bool find_highest, uint64_t *result_keys)
901 {
902 int r = 0, count = 0, level;
903 struct ro_spine spine;
904
905 init_ro_spine(&spine, info);
906 for (level = 0; level < info->levels; level++) {
907 r = find_key(&spine, root, find_highest, result_keys + level,
908 level == info->levels - 1 ? NULL : &root);
909 if (r == -ENODATA) {
910 r = 0;
911 break;
912
913 } else if (r)
914 break;
915
916 count++;
917 }
918 exit_ro_spine(&spine);
919
920 return r ? r : count;
921 }
922
dm_btree_find_highest_key(struct dm_btree_info * info,dm_block_t root,uint64_t * result_keys)923 int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
924 uint64_t *result_keys)
925 {
926 return dm_btree_find_key(info, root, true, result_keys);
927 }
928 EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
929
dm_btree_find_lowest_key(struct dm_btree_info * info,dm_block_t root,uint64_t * result_keys)930 int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
931 uint64_t *result_keys)
932 {
933 return dm_btree_find_key(info, root, false, result_keys);
934 }
935 EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
936
937 /*----------------------------------------------------------------*/
938
939 /*
940 * FIXME: We shouldn't use a recursive algorithm when we have limited stack
941 * space. Also this only works for single level trees.
942 */
walk_node(struct dm_btree_info * info,dm_block_t block,int (* fn)(void * context,uint64_t * keys,void * leaf),void * context)943 static int walk_node(struct dm_btree_info *info, dm_block_t block,
944 int (*fn)(void *context, uint64_t *keys, void *leaf),
945 void *context)
946 {
947 int r;
948 unsigned i, nr;
949 struct dm_block *node;
950 struct btree_node *n;
951 uint64_t keys;
952
953 r = bn_read_lock(info, block, &node);
954 if (r)
955 return r;
956
957 n = dm_block_data(node);
958
959 nr = le32_to_cpu(n->header.nr_entries);
960 for (i = 0; i < nr; i++) {
961 if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
962 r = walk_node(info, value64(n, i), fn, context);
963 if (r)
964 goto out;
965 } else {
966 keys = le64_to_cpu(*key_ptr(n, i));
967 r = fn(context, &keys, value_ptr(n, i));
968 if (r)
969 goto out;
970 }
971 }
972
973 out:
974 dm_tm_unlock(info->tm, node);
975 return r;
976 }
977
dm_btree_walk(struct dm_btree_info * info,dm_block_t root,int (* fn)(void * context,uint64_t * keys,void * leaf),void * context)978 int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
979 int (*fn)(void *context, uint64_t *keys, void *leaf),
980 void *context)
981 {
982 BUG_ON(info->levels > 1);
983 return walk_node(info, root, fn, context);
984 }
985 EXPORT_SYMBOL_GPL(dm_btree_walk);
986
987 /*----------------------------------------------------------------*/
988
prefetch_values(struct dm_btree_cursor * c)989 static void prefetch_values(struct dm_btree_cursor *c)
990 {
991 unsigned i, nr;
992 __le64 value_le;
993 struct cursor_node *n = c->nodes + c->depth - 1;
994 struct btree_node *bn = dm_block_data(n->b);
995 struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
996
997 BUG_ON(c->info->value_type.size != sizeof(value_le));
998
999 nr = le32_to_cpu(bn->header.nr_entries);
1000 for (i = 0; i < nr; i++) {
1001 memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
1002 dm_bm_prefetch(bm, le64_to_cpu(value_le));
1003 }
1004 }
1005
leaf_node(struct dm_btree_cursor * c)1006 static bool leaf_node(struct dm_btree_cursor *c)
1007 {
1008 struct cursor_node *n = c->nodes + c->depth - 1;
1009 struct btree_node *bn = dm_block_data(n->b);
1010
1011 return le32_to_cpu(bn->header.flags) & LEAF_NODE;
1012 }
1013
push_node(struct dm_btree_cursor * c,dm_block_t b)1014 static int push_node(struct dm_btree_cursor *c, dm_block_t b)
1015 {
1016 int r;
1017 struct cursor_node *n = c->nodes + c->depth;
1018
1019 if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
1020 DMERR("couldn't push cursor node, stack depth too high");
1021 return -EINVAL;
1022 }
1023
1024 r = bn_read_lock(c->info, b, &n->b);
1025 if (r)
1026 return r;
1027
1028 n->index = 0;
1029 c->depth++;
1030
1031 if (c->prefetch_leaves || !leaf_node(c))
1032 prefetch_values(c);
1033
1034 return 0;
1035 }
1036
pop_node(struct dm_btree_cursor * c)1037 static void pop_node(struct dm_btree_cursor *c)
1038 {
1039 c->depth--;
1040 unlock_block(c->info, c->nodes[c->depth].b);
1041 }
1042
inc_or_backtrack(struct dm_btree_cursor * c)1043 static int inc_or_backtrack(struct dm_btree_cursor *c)
1044 {
1045 struct cursor_node *n;
1046 struct btree_node *bn;
1047
1048 for (;;) {
1049 if (!c->depth)
1050 return -ENODATA;
1051
1052 n = c->nodes + c->depth - 1;
1053 bn = dm_block_data(n->b);
1054
1055 n->index++;
1056 if (n->index < le32_to_cpu(bn->header.nr_entries))
1057 break;
1058
1059 pop_node(c);
1060 }
1061
1062 return 0;
1063 }
1064
find_leaf(struct dm_btree_cursor * c)1065 static int find_leaf(struct dm_btree_cursor *c)
1066 {
1067 int r = 0;
1068 struct cursor_node *n;
1069 struct btree_node *bn;
1070 __le64 value_le;
1071
1072 for (;;) {
1073 n = c->nodes + c->depth - 1;
1074 bn = dm_block_data(n->b);
1075
1076 if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
1077 break;
1078
1079 memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
1080 r = push_node(c, le64_to_cpu(value_le));
1081 if (r) {
1082 DMERR("push_node failed");
1083 break;
1084 }
1085 }
1086
1087 if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
1088 return -ENODATA;
1089
1090 return r;
1091 }
1092
dm_btree_cursor_begin(struct dm_btree_info * info,dm_block_t root,bool prefetch_leaves,struct dm_btree_cursor * c)1093 int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
1094 bool prefetch_leaves, struct dm_btree_cursor *c)
1095 {
1096 int r;
1097
1098 c->info = info;
1099 c->root = root;
1100 c->depth = 0;
1101 c->prefetch_leaves = prefetch_leaves;
1102
1103 r = push_node(c, root);
1104 if (r)
1105 return r;
1106
1107 return find_leaf(c);
1108 }
1109 EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
1110
dm_btree_cursor_end(struct dm_btree_cursor * c)1111 void dm_btree_cursor_end(struct dm_btree_cursor *c)
1112 {
1113 while (c->depth)
1114 pop_node(c);
1115 }
1116 EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
1117
dm_btree_cursor_next(struct dm_btree_cursor * c)1118 int dm_btree_cursor_next(struct dm_btree_cursor *c)
1119 {
1120 int r = inc_or_backtrack(c);
1121 if (!r) {
1122 r = find_leaf(c);
1123 if (r)
1124 DMERR("find_leaf failed");
1125 }
1126
1127 return r;
1128 }
1129 EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
1130
dm_btree_cursor_get_value(struct dm_btree_cursor * c,uint64_t * key,void * value_le)1131 int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
1132 {
1133 if (c->depth) {
1134 struct cursor_node *n = c->nodes + c->depth - 1;
1135 struct btree_node *bn = dm_block_data(n->b);
1136
1137 if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
1138 return -EINVAL;
1139
1140 *key = le64_to_cpu(*key_ptr(bn, n->index));
1141 memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
1142 return 0;
1143
1144 } else
1145 return -ENODATA;
1146 }
1147 EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);
1148