dm-btree.c 25 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148
  1. /*
  2. * Copyright (C) 2011 Red Hat, Inc.
  3. *
  4. * This file is released under the GPL.
  5. */
  6. #include "dm-btree-internal.h"
  7. #include "dm-space-map.h"
  8. #include "dm-transaction-manager.h"
  9. #include <linux/export.h>
  10. #include <linux/device-mapper.h>
  11. #define DM_MSG_PREFIX "btree"
  12. /*----------------------------------------------------------------
  13. * Array manipulation
  14. *--------------------------------------------------------------*/
  15. static void memcpy_disk(void *dest, const void *src, size_t len)
  16. __dm_written_to_disk(src)
  17. {
  18. memcpy(dest, src, len);
  19. __dm_unbless_for_disk(src);
  20. }
  21. static void array_insert(void *base, size_t elt_size, unsigned nr_elts,
  22. unsigned index, void *elt)
  23. __dm_written_to_disk(elt)
  24. {
  25. if (index < nr_elts)
  26. memmove(base + (elt_size * (index + 1)),
  27. base + (elt_size * index),
  28. (nr_elts - index) * elt_size);
  29. memcpy_disk(base + (elt_size * index), elt, elt_size);
  30. }
  31. /*----------------------------------------------------------------*/
  32. /* makes the assumption that no two keys are the same. */
  33. static int bsearch(struct btree_node *n, uint64_t key, int want_hi)
  34. {
  35. int lo = -1, hi = le32_to_cpu(n->header.nr_entries);
  36. while (hi - lo > 1) {
  37. int mid = lo + ((hi - lo) / 2);
  38. uint64_t mid_key = le64_to_cpu(n->keys[mid]);
  39. if (mid_key == key)
  40. return mid;
  41. if (mid_key < key)
  42. lo = mid;
  43. else
  44. hi = mid;
  45. }
  46. return want_hi ? hi : lo;
  47. }
  48. int lower_bound(struct btree_node *n, uint64_t key)
  49. {
  50. return bsearch(n, key, 0);
  51. }
  52. static int upper_bound(struct btree_node *n, uint64_t key)
  53. {
  54. return bsearch(n, key, 1);
  55. }
  56. void inc_children(struct dm_transaction_manager *tm, struct btree_node *n,
  57. struct dm_btree_value_type *vt)
  58. {
  59. unsigned i;
  60. uint32_t nr_entries = le32_to_cpu(n->header.nr_entries);
  61. if (le32_to_cpu(n->header.flags) & INTERNAL_NODE)
  62. for (i = 0; i < nr_entries; i++)
  63. dm_tm_inc(tm, value64(n, i));
  64. else if (vt->inc)
  65. for (i = 0; i < nr_entries; i++)
  66. vt->inc(vt->context, value_ptr(n, i));
  67. }
  68. static int insert_at(size_t value_size, struct btree_node *node, unsigned index,
  69. uint64_t key, void *value)
  70. __dm_written_to_disk(value)
  71. {
  72. uint32_t nr_entries = le32_to_cpu(node->header.nr_entries);
  73. __le64 key_le = cpu_to_le64(key);
  74. if (index > nr_entries ||
  75. index >= le32_to_cpu(node->header.max_entries)) {
  76. DMERR("too many entries in btree node for insert");
  77. __dm_unbless_for_disk(value);
  78. return -ENOMEM;
  79. }
  80. __dm_bless_for_disk(&key_le);
  81. array_insert(node->keys, sizeof(*node->keys), nr_entries, index, &key_le);
  82. array_insert(value_base(node), value_size, nr_entries, index, value);
  83. node->header.nr_entries = cpu_to_le32(nr_entries + 1);
  84. return 0;
  85. }
  86. /*----------------------------------------------------------------*/
  87. /*
  88. * We want 3n entries (for some n). This works more nicely for repeated
  89. * insert remove loops than (2n + 1).
  90. */
  91. static uint32_t calc_max_entries(size_t value_size, size_t block_size)
  92. {
  93. uint32_t total, n;
  94. size_t elt_size = sizeof(uint64_t) + value_size; /* key + value */
  95. block_size -= sizeof(struct node_header);
  96. total = block_size / elt_size;
  97. n = total / 3; /* rounds down */
  98. return 3 * n;
  99. }
  100. int dm_btree_empty(struct dm_btree_info *info, dm_block_t *root)
  101. {
  102. int r;
  103. struct dm_block *b;
  104. struct btree_node *n;
  105. size_t block_size;
  106. uint32_t max_entries;
  107. r = new_block(info, &b);
  108. if (r < 0)
  109. return r;
  110. block_size = dm_bm_block_size(dm_tm_get_bm(info->tm));
  111. max_entries = calc_max_entries(info->value_type.size, block_size);
  112. n = dm_block_data(b);
  113. memset(n, 0, block_size);
  114. n->header.flags = cpu_to_le32(LEAF_NODE);
  115. n->header.nr_entries = cpu_to_le32(0);
  116. n->header.max_entries = cpu_to_le32(max_entries);
  117. n->header.value_size = cpu_to_le32(info->value_type.size);
  118. *root = dm_block_location(b);
  119. unlock_block(info, b);
  120. return 0;
  121. }
  122. EXPORT_SYMBOL_GPL(dm_btree_empty);
  123. /*----------------------------------------------------------------*/
  124. /*
  125. * Deletion uses a recursive algorithm, since we have limited stack space
  126. * we explicitly manage our own stack on the heap.
  127. */
  128. #define MAX_SPINE_DEPTH 64
  129. struct frame {
  130. struct dm_block *b;
  131. struct btree_node *n;
  132. unsigned level;
  133. unsigned nr_children;
  134. unsigned current_child;
  135. };
  136. struct del_stack {
  137. struct dm_btree_info *info;
  138. struct dm_transaction_manager *tm;
  139. int top;
  140. struct frame spine[MAX_SPINE_DEPTH];
  141. };
  142. static int top_frame(struct del_stack *s, struct frame **f)
  143. {
  144. if (s->top < 0) {
  145. DMERR("btree deletion stack empty");
  146. return -EINVAL;
  147. }
  148. *f = s->spine + s->top;
  149. return 0;
  150. }
  151. static int unprocessed_frames(struct del_stack *s)
  152. {
  153. return s->top >= 0;
  154. }
  155. static void prefetch_children(struct del_stack *s, struct frame *f)
  156. {
  157. unsigned i;
  158. struct dm_block_manager *bm = dm_tm_get_bm(s->tm);
  159. for (i = 0; i < f->nr_children; i++)
  160. dm_bm_prefetch(bm, value64(f->n, i));
  161. }
  162. static bool is_internal_level(struct dm_btree_info *info, struct frame *f)
  163. {
  164. return f->level < (info->levels - 1);
  165. }
  166. static int push_frame(struct del_stack *s, dm_block_t b, unsigned level)
  167. {
  168. int r;
  169. uint32_t ref_count;
  170. if (s->top >= MAX_SPINE_DEPTH - 1) {
  171. DMERR("btree deletion stack out of memory");
  172. return -ENOMEM;
  173. }
  174. r = dm_tm_ref(s->tm, b, &ref_count);
  175. if (r)
  176. return r;
  177. if (ref_count > 1)
  178. /*
  179. * This is a shared node, so we can just decrement it's
  180. * reference counter and leave the children.
  181. */
  182. dm_tm_dec(s->tm, b);
  183. else {
  184. uint32_t flags;
  185. struct frame *f = s->spine + ++s->top;
  186. r = dm_tm_read_lock(s->tm, b, &btree_node_validator, &f->b);
  187. if (r) {
  188. s->top--;
  189. return r;
  190. }
  191. f->n = dm_block_data(f->b);
  192. f->level = level;
  193. f->nr_children = le32_to_cpu(f->n->header.nr_entries);
  194. f->current_child = 0;
  195. flags = le32_to_cpu(f->n->header.flags);
  196. if (flags & INTERNAL_NODE || is_internal_level(s->info, f))
  197. prefetch_children(s, f);
  198. }
  199. return 0;
  200. }
  201. static void pop_frame(struct del_stack *s)
  202. {
  203. struct frame *f = s->spine + s->top--;
  204. dm_tm_dec(s->tm, dm_block_location(f->b));
  205. dm_tm_unlock(s->tm, f->b);
  206. }
  207. static void unlock_all_frames(struct del_stack *s)
  208. {
  209. struct frame *f;
  210. while (unprocessed_frames(s)) {
  211. f = s->spine + s->top--;
  212. dm_tm_unlock(s->tm, f->b);
  213. }
  214. }
  215. int dm_btree_del(struct dm_btree_info *info, dm_block_t root)
  216. {
  217. int r;
  218. struct del_stack *s;
  219. s = kmalloc(sizeof(*s), GFP_NOIO);
  220. if (!s)
  221. return -ENOMEM;
  222. s->info = info;
  223. s->tm = info->tm;
  224. s->top = -1;
  225. r = push_frame(s, root, 0);
  226. if (r)
  227. goto out;
  228. while (unprocessed_frames(s)) {
  229. uint32_t flags;
  230. struct frame *f;
  231. dm_block_t b;
  232. r = top_frame(s, &f);
  233. if (r)
  234. goto out;
  235. if (f->current_child >= f->nr_children) {
  236. pop_frame(s);
  237. continue;
  238. }
  239. flags = le32_to_cpu(f->n->header.flags);
  240. if (flags & INTERNAL_NODE) {
  241. b = value64(f->n, f->current_child);
  242. f->current_child++;
  243. r = push_frame(s, b, f->level);
  244. if (r)
  245. goto out;
  246. } else if (is_internal_level(info, f)) {
  247. b = value64(f->n, f->current_child);
  248. f->current_child++;
  249. r = push_frame(s, b, f->level + 1);
  250. if (r)
  251. goto out;
  252. } else {
  253. if (info->value_type.dec) {
  254. unsigned i;
  255. for (i = 0; i < f->nr_children; i++)
  256. info->value_type.dec(info->value_type.context,
  257. value_ptr(f->n, i));
  258. }
  259. pop_frame(s);
  260. }
  261. }
  262. out:
  263. if (r) {
  264. /* cleanup all frames of del_stack */
  265. unlock_all_frames(s);
  266. }
  267. kfree(s);
  268. return r;
  269. }
  270. EXPORT_SYMBOL_GPL(dm_btree_del);
  271. /*----------------------------------------------------------------*/
  272. static int btree_lookup_raw(struct ro_spine *s, dm_block_t block, uint64_t key,
  273. int (*search_fn)(struct btree_node *, uint64_t),
  274. uint64_t *result_key, void *v, size_t value_size)
  275. {
  276. int i, r;
  277. uint32_t flags, nr_entries;
  278. do {
  279. r = ro_step(s, block);
  280. if (r < 0)
  281. return r;
  282. i = search_fn(ro_node(s), key);
  283. flags = le32_to_cpu(ro_node(s)->header.flags);
  284. nr_entries = le32_to_cpu(ro_node(s)->header.nr_entries);
  285. if (i < 0 || i >= nr_entries)
  286. return -ENODATA;
  287. if (flags & INTERNAL_NODE)
  288. block = value64(ro_node(s), i);
  289. } while (!(flags & LEAF_NODE));
  290. *result_key = le64_to_cpu(ro_node(s)->keys[i]);
  291. memcpy(v, value_ptr(ro_node(s), i), value_size);
  292. return 0;
  293. }
  294. int dm_btree_lookup(struct dm_btree_info *info, dm_block_t root,
  295. uint64_t *keys, void *value_le)
  296. {
  297. unsigned level, last_level = info->levels - 1;
  298. int r = -ENODATA;
  299. uint64_t rkey;
  300. __le64 internal_value_le;
  301. struct ro_spine spine;
  302. init_ro_spine(&spine, info);
  303. for (level = 0; level < info->levels; level++) {
  304. size_t size;
  305. void *value_p;
  306. if (level == last_level) {
  307. value_p = value_le;
  308. size = info->value_type.size;
  309. } else {
  310. value_p = &internal_value_le;
  311. size = sizeof(uint64_t);
  312. }
  313. r = btree_lookup_raw(&spine, root, keys[level],
  314. lower_bound, &rkey,
  315. value_p, size);
  316. if (!r) {
  317. if (rkey != keys[level]) {
  318. exit_ro_spine(&spine);
  319. return -ENODATA;
  320. }
  321. } else {
  322. exit_ro_spine(&spine);
  323. return r;
  324. }
  325. root = le64_to_cpu(internal_value_le);
  326. }
  327. exit_ro_spine(&spine);
  328. return r;
  329. }
  330. EXPORT_SYMBOL_GPL(dm_btree_lookup);
  331. static int dm_btree_lookup_next_single(struct dm_btree_info *info, dm_block_t root,
  332. uint64_t key, uint64_t *rkey, void *value_le)
  333. {
  334. int r, i;
  335. uint32_t flags, nr_entries;
  336. struct dm_block *node;
  337. struct btree_node *n;
  338. r = bn_read_lock(info, root, &node);
  339. if (r)
  340. return r;
  341. n = dm_block_data(node);
  342. flags = le32_to_cpu(n->header.flags);
  343. nr_entries = le32_to_cpu(n->header.nr_entries);
  344. if (flags & INTERNAL_NODE) {
  345. i = lower_bound(n, key);
  346. if (i < 0) {
  347. /*
  348. * avoid early -ENODATA return when all entries are
  349. * higher than the search @key.
  350. */
  351. i = 0;
  352. }
  353. if (i >= nr_entries) {
  354. r = -ENODATA;
  355. goto out;
  356. }
  357. r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
  358. if (r == -ENODATA && i < (nr_entries - 1)) {
  359. i++;
  360. r = dm_btree_lookup_next_single(info, value64(n, i), key, rkey, value_le);
  361. }
  362. } else {
  363. i = upper_bound(n, key);
  364. if (i < 0 || i >= nr_entries) {
  365. r = -ENODATA;
  366. goto out;
  367. }
  368. *rkey = le64_to_cpu(n->keys[i]);
  369. memcpy(value_le, value_ptr(n, i), info->value_type.size);
  370. }
  371. out:
  372. dm_tm_unlock(info->tm, node);
  373. return r;
  374. }
  375. int dm_btree_lookup_next(struct dm_btree_info *info, dm_block_t root,
  376. uint64_t *keys, uint64_t *rkey, void *value_le)
  377. {
  378. unsigned level;
  379. int r = -ENODATA;
  380. __le64 internal_value_le;
  381. struct ro_spine spine;
  382. init_ro_spine(&spine, info);
  383. for (level = 0; level < info->levels - 1u; level++) {
  384. r = btree_lookup_raw(&spine, root, keys[level],
  385. lower_bound, rkey,
  386. &internal_value_le, sizeof(uint64_t));
  387. if (r)
  388. goto out;
  389. if (*rkey != keys[level]) {
  390. r = -ENODATA;
  391. goto out;
  392. }
  393. root = le64_to_cpu(internal_value_le);
  394. }
  395. r = dm_btree_lookup_next_single(info, root, keys[level], rkey, value_le);
  396. out:
  397. exit_ro_spine(&spine);
  398. return r;
  399. }
  400. EXPORT_SYMBOL_GPL(dm_btree_lookup_next);
  401. /*
  402. * Splits a node by creating a sibling node and shifting half the nodes
  403. * contents across. Assumes there is a parent node, and it has room for
  404. * another child.
  405. *
  406. * Before:
  407. * +--------+
  408. * | Parent |
  409. * +--------+
  410. * |
  411. * v
  412. * +----------+
  413. * | A ++++++ |
  414. * +----------+
  415. *
  416. *
  417. * After:
  418. * +--------+
  419. * | Parent |
  420. * +--------+
  421. * | |
  422. * v +------+
  423. * +---------+ |
  424. * | A* +++ | v
  425. * +---------+ +-------+
  426. * | B +++ |
  427. * +-------+
  428. *
  429. * Where A* is a shadow of A.
  430. */
  431. static int btree_split_sibling(struct shadow_spine *s, unsigned parent_index,
  432. uint64_t key)
  433. {
  434. int r;
  435. size_t size;
  436. unsigned nr_left, nr_right;
  437. struct dm_block *left, *right, *parent;
  438. struct btree_node *ln, *rn, *pn;
  439. __le64 location;
  440. left = shadow_current(s);
  441. r = new_block(s->info, &right);
  442. if (r < 0)
  443. return r;
  444. ln = dm_block_data(left);
  445. rn = dm_block_data(right);
  446. nr_left = le32_to_cpu(ln->header.nr_entries) / 2;
  447. nr_right = le32_to_cpu(ln->header.nr_entries) - nr_left;
  448. ln->header.nr_entries = cpu_to_le32(nr_left);
  449. rn->header.flags = ln->header.flags;
  450. rn->header.nr_entries = cpu_to_le32(nr_right);
  451. rn->header.max_entries = ln->header.max_entries;
  452. rn->header.value_size = ln->header.value_size;
  453. memcpy(rn->keys, ln->keys + nr_left, nr_right * sizeof(rn->keys[0]));
  454. size = le32_to_cpu(ln->header.flags) & INTERNAL_NODE ?
  455. sizeof(uint64_t) : s->info->value_type.size;
  456. memcpy(value_ptr(rn, 0), value_ptr(ln, nr_left),
  457. size * nr_right);
  458. /*
  459. * Patch up the parent
  460. */
  461. parent = shadow_parent(s);
  462. pn = dm_block_data(parent);
  463. location = cpu_to_le64(dm_block_location(left));
  464. __dm_bless_for_disk(&location);
  465. memcpy_disk(value_ptr(pn, parent_index),
  466. &location, sizeof(__le64));
  467. location = cpu_to_le64(dm_block_location(right));
  468. __dm_bless_for_disk(&location);
  469. r = insert_at(sizeof(__le64), pn, parent_index + 1,
  470. le64_to_cpu(rn->keys[0]), &location);
  471. if (r) {
  472. unlock_block(s->info, right);
  473. return r;
  474. }
  475. if (key < le64_to_cpu(rn->keys[0])) {
  476. unlock_block(s->info, right);
  477. s->nodes[1] = left;
  478. } else {
  479. unlock_block(s->info, left);
  480. s->nodes[1] = right;
  481. }
  482. return 0;
  483. }
  484. /*
  485. * Splits a node by creating two new children beneath the given node.
  486. *
  487. * Before:
  488. * +----------+
  489. * | A ++++++ |
  490. * +----------+
  491. *
  492. *
  493. * After:
  494. * +------------+
  495. * | A (shadow) |
  496. * +------------+
  497. * | |
  498. * +------+ +----+
  499. * | |
  500. * v v
  501. * +-------+ +-------+
  502. * | B +++ | | C +++ |
  503. * +-------+ +-------+
  504. */
  505. static int btree_split_beneath(struct shadow_spine *s, uint64_t key)
  506. {
  507. int r;
  508. size_t size;
  509. unsigned nr_left, nr_right;
  510. struct dm_block *left, *right, *new_parent;
  511. struct btree_node *pn, *ln, *rn;
  512. __le64 val;
  513. new_parent = shadow_current(s);
  514. r = new_block(s->info, &left);
  515. if (r < 0)
  516. return r;
  517. r = new_block(s->info, &right);
  518. if (r < 0) {
  519. unlock_block(s->info, left);
  520. return r;
  521. }
  522. pn = dm_block_data(new_parent);
  523. ln = dm_block_data(left);
  524. rn = dm_block_data(right);
  525. nr_left = le32_to_cpu(pn->header.nr_entries) / 2;
  526. nr_right = le32_to_cpu(pn->header.nr_entries) - nr_left;
  527. ln->header.flags = pn->header.flags;
  528. ln->header.nr_entries = cpu_to_le32(nr_left);
  529. ln->header.max_entries = pn->header.max_entries;
  530. ln->header.value_size = pn->header.value_size;
  531. rn->header.flags = pn->header.flags;
  532. rn->header.nr_entries = cpu_to_le32(nr_right);
  533. rn->header.max_entries = pn->header.max_entries;
  534. rn->header.value_size = pn->header.value_size;
  535. memcpy(ln->keys, pn->keys, nr_left * sizeof(pn->keys[0]));
  536. memcpy(rn->keys, pn->keys + nr_left, nr_right * sizeof(pn->keys[0]));
  537. size = le32_to_cpu(pn->header.flags) & INTERNAL_NODE ?
  538. sizeof(__le64) : s->info->value_type.size;
  539. memcpy(value_ptr(ln, 0), value_ptr(pn, 0), nr_left * size);
  540. memcpy(value_ptr(rn, 0), value_ptr(pn, nr_left),
  541. nr_right * size);
  542. /* new_parent should just point to l and r now */
  543. pn->header.flags = cpu_to_le32(INTERNAL_NODE);
  544. pn->header.nr_entries = cpu_to_le32(2);
  545. pn->header.max_entries = cpu_to_le32(
  546. calc_max_entries(sizeof(__le64),
  547. dm_bm_block_size(
  548. dm_tm_get_bm(s->info->tm))));
  549. pn->header.value_size = cpu_to_le32(sizeof(__le64));
  550. val = cpu_to_le64(dm_block_location(left));
  551. __dm_bless_for_disk(&val);
  552. pn->keys[0] = ln->keys[0];
  553. memcpy_disk(value_ptr(pn, 0), &val, sizeof(__le64));
  554. val = cpu_to_le64(dm_block_location(right));
  555. __dm_bless_for_disk(&val);
  556. pn->keys[1] = rn->keys[0];
  557. memcpy_disk(value_ptr(pn, 1), &val, sizeof(__le64));
  558. unlock_block(s->info, left);
  559. unlock_block(s->info, right);
  560. return 0;
  561. }
  562. static int btree_insert_raw(struct shadow_spine *s, dm_block_t root,
  563. struct dm_btree_value_type *vt,
  564. uint64_t key, unsigned *index)
  565. {
  566. int r, i = *index, top = 1;
  567. struct btree_node *node;
  568. for (;;) {
  569. r = shadow_step(s, root, vt);
  570. if (r < 0)
  571. return r;
  572. node = dm_block_data(shadow_current(s));
  573. /*
  574. * We have to patch up the parent node, ugly, but I don't
  575. * see a way to do this automatically as part of the spine
  576. * op.
  577. */
  578. if (shadow_has_parent(s) && i >= 0) { /* FIXME: second clause unness. */
  579. __le64 location = cpu_to_le64(dm_block_location(shadow_current(s)));
  580. __dm_bless_for_disk(&location);
  581. memcpy_disk(value_ptr(dm_block_data(shadow_parent(s)), i),
  582. &location, sizeof(__le64));
  583. }
  584. node = dm_block_data(shadow_current(s));
  585. if (node->header.nr_entries == node->header.max_entries) {
  586. if (top)
  587. r = btree_split_beneath(s, key);
  588. else
  589. r = btree_split_sibling(s, i, key);
  590. if (r < 0)
  591. return r;
  592. }
  593. node = dm_block_data(shadow_current(s));
  594. i = lower_bound(node, key);
  595. if (le32_to_cpu(node->header.flags) & LEAF_NODE)
  596. break;
  597. if (i < 0) {
  598. /* change the bounds on the lowest key */
  599. node->keys[0] = cpu_to_le64(key);
  600. i = 0;
  601. }
  602. root = value64(node, i);
  603. top = 0;
  604. }
  605. if (i < 0 || le64_to_cpu(node->keys[i]) != key)
  606. i++;
  607. *index = i;
  608. return 0;
  609. }
  610. static bool need_insert(struct btree_node *node, uint64_t *keys,
  611. unsigned level, unsigned index)
  612. {
  613. return ((index >= le32_to_cpu(node->header.nr_entries)) ||
  614. (le64_to_cpu(node->keys[index]) != keys[level]));
  615. }
  616. static int insert(struct dm_btree_info *info, dm_block_t root,
  617. uint64_t *keys, void *value, dm_block_t *new_root,
  618. int *inserted)
  619. __dm_written_to_disk(value)
  620. {
  621. int r;
  622. unsigned level, index = -1, last_level = info->levels - 1;
  623. dm_block_t block = root;
  624. struct shadow_spine spine;
  625. struct btree_node *n;
  626. struct dm_btree_value_type le64_type;
  627. init_le64_type(info->tm, &le64_type);
  628. init_shadow_spine(&spine, info);
  629. for (level = 0; level < (info->levels - 1); level++) {
  630. r = btree_insert_raw(&spine, block, &le64_type, keys[level], &index);
  631. if (r < 0)
  632. goto bad;
  633. n = dm_block_data(shadow_current(&spine));
  634. if (need_insert(n, keys, level, index)) {
  635. dm_block_t new_tree;
  636. __le64 new_le;
  637. r = dm_btree_empty(info, &new_tree);
  638. if (r < 0)
  639. goto bad;
  640. new_le = cpu_to_le64(new_tree);
  641. __dm_bless_for_disk(&new_le);
  642. r = insert_at(sizeof(uint64_t), n, index,
  643. keys[level], &new_le);
  644. if (r)
  645. goto bad;
  646. }
  647. if (level < last_level)
  648. block = value64(n, index);
  649. }
  650. r = btree_insert_raw(&spine, block, &info->value_type,
  651. keys[level], &index);
  652. if (r < 0)
  653. goto bad;
  654. n = dm_block_data(shadow_current(&spine));
  655. if (need_insert(n, keys, level, index)) {
  656. if (inserted)
  657. *inserted = 1;
  658. r = insert_at(info->value_type.size, n, index,
  659. keys[level], value);
  660. if (r)
  661. goto bad_unblessed;
  662. } else {
  663. if (inserted)
  664. *inserted = 0;
  665. if (info->value_type.dec &&
  666. (!info->value_type.equal ||
  667. !info->value_type.equal(
  668. info->value_type.context,
  669. value_ptr(n, index),
  670. value))) {
  671. info->value_type.dec(info->value_type.context,
  672. value_ptr(n, index));
  673. }
  674. memcpy_disk(value_ptr(n, index),
  675. value, info->value_type.size);
  676. }
  677. *new_root = shadow_root(&spine);
  678. exit_shadow_spine(&spine);
  679. return 0;
  680. bad:
  681. __dm_unbless_for_disk(value);
  682. bad_unblessed:
  683. exit_shadow_spine(&spine);
  684. return r;
  685. }
  686. int dm_btree_insert(struct dm_btree_info *info, dm_block_t root,
  687. uint64_t *keys, void *value, dm_block_t *new_root)
  688. __dm_written_to_disk(value)
  689. {
  690. return insert(info, root, keys, value, new_root, NULL);
  691. }
  692. EXPORT_SYMBOL_GPL(dm_btree_insert);
  693. int dm_btree_insert_notify(struct dm_btree_info *info, dm_block_t root,
  694. uint64_t *keys, void *value, dm_block_t *new_root,
  695. int *inserted)
  696. __dm_written_to_disk(value)
  697. {
  698. return insert(info, root, keys, value, new_root, inserted);
  699. }
  700. EXPORT_SYMBOL_GPL(dm_btree_insert_notify);
  701. /*----------------------------------------------------------------*/
  702. static int find_key(struct ro_spine *s, dm_block_t block, bool find_highest,
  703. uint64_t *result_key, dm_block_t *next_block)
  704. {
  705. int i, r;
  706. uint32_t flags;
  707. do {
  708. r = ro_step(s, block);
  709. if (r < 0)
  710. return r;
  711. flags = le32_to_cpu(ro_node(s)->header.flags);
  712. i = le32_to_cpu(ro_node(s)->header.nr_entries);
  713. if (!i)
  714. return -ENODATA;
  715. else
  716. i--;
  717. if (find_highest)
  718. *result_key = le64_to_cpu(ro_node(s)->keys[i]);
  719. else
  720. *result_key = le64_to_cpu(ro_node(s)->keys[0]);
  721. if (next_block || flags & INTERNAL_NODE) {
  722. if (find_highest)
  723. block = value64(ro_node(s), i);
  724. else
  725. block = value64(ro_node(s), 0);
  726. }
  727. } while (flags & INTERNAL_NODE);
  728. if (next_block)
  729. *next_block = block;
  730. return 0;
  731. }
  732. static int dm_btree_find_key(struct dm_btree_info *info, dm_block_t root,
  733. bool find_highest, uint64_t *result_keys)
  734. {
  735. int r = 0, count = 0, level;
  736. struct ro_spine spine;
  737. init_ro_spine(&spine, info);
  738. for (level = 0; level < info->levels; level++) {
  739. r = find_key(&spine, root, find_highest, result_keys + level,
  740. level == info->levels - 1 ? NULL : &root);
  741. if (r == -ENODATA) {
  742. r = 0;
  743. break;
  744. } else if (r)
  745. break;
  746. count++;
  747. }
  748. exit_ro_spine(&spine);
  749. return r ? r : count;
  750. }
  751. int dm_btree_find_highest_key(struct dm_btree_info *info, dm_block_t root,
  752. uint64_t *result_keys)
  753. {
  754. return dm_btree_find_key(info, root, true, result_keys);
  755. }
  756. EXPORT_SYMBOL_GPL(dm_btree_find_highest_key);
  757. int dm_btree_find_lowest_key(struct dm_btree_info *info, dm_block_t root,
  758. uint64_t *result_keys)
  759. {
  760. return dm_btree_find_key(info, root, false, result_keys);
  761. }
  762. EXPORT_SYMBOL_GPL(dm_btree_find_lowest_key);
  763. /*----------------------------------------------------------------*/
  764. /*
  765. * FIXME: We shouldn't use a recursive algorithm when we have limited stack
  766. * space. Also this only works for single level trees.
  767. */
  768. static int walk_node(struct dm_btree_info *info, dm_block_t block,
  769. int (*fn)(void *context, uint64_t *keys, void *leaf),
  770. void *context)
  771. {
  772. int r;
  773. unsigned i, nr;
  774. struct dm_block *node;
  775. struct btree_node *n;
  776. uint64_t keys;
  777. r = bn_read_lock(info, block, &node);
  778. if (r)
  779. return r;
  780. n = dm_block_data(node);
  781. nr = le32_to_cpu(n->header.nr_entries);
  782. for (i = 0; i < nr; i++) {
  783. if (le32_to_cpu(n->header.flags) & INTERNAL_NODE) {
  784. r = walk_node(info, value64(n, i), fn, context);
  785. if (r)
  786. goto out;
  787. } else {
  788. keys = le64_to_cpu(*key_ptr(n, i));
  789. r = fn(context, &keys, value_ptr(n, i));
  790. if (r)
  791. goto out;
  792. }
  793. }
  794. out:
  795. dm_tm_unlock(info->tm, node);
  796. return r;
  797. }
  798. int dm_btree_walk(struct dm_btree_info *info, dm_block_t root,
  799. int (*fn)(void *context, uint64_t *keys, void *leaf),
  800. void *context)
  801. {
  802. BUG_ON(info->levels > 1);
  803. return walk_node(info, root, fn, context);
  804. }
  805. EXPORT_SYMBOL_GPL(dm_btree_walk);
  806. /*----------------------------------------------------------------*/
  807. static void prefetch_values(struct dm_btree_cursor *c)
  808. {
  809. unsigned i, nr;
  810. __le64 value_le;
  811. struct cursor_node *n = c->nodes + c->depth - 1;
  812. struct btree_node *bn = dm_block_data(n->b);
  813. struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
  814. BUG_ON(c->info->value_type.size != sizeof(value_le));
  815. nr = le32_to_cpu(bn->header.nr_entries);
  816. for (i = 0; i < nr; i++) {
  817. memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
  818. dm_bm_prefetch(bm, le64_to_cpu(value_le));
  819. }
  820. }
  821. static bool leaf_node(struct dm_btree_cursor *c)
  822. {
  823. struct cursor_node *n = c->nodes + c->depth - 1;
  824. struct btree_node *bn = dm_block_data(n->b);
  825. return le32_to_cpu(bn->header.flags) & LEAF_NODE;
  826. }
  827. static int push_node(struct dm_btree_cursor *c, dm_block_t b)
  828. {
  829. int r;
  830. struct cursor_node *n = c->nodes + c->depth;
  831. if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
  832. DMERR("couldn't push cursor node, stack depth too high");
  833. return -EINVAL;
  834. }
  835. r = bn_read_lock(c->info, b, &n->b);
  836. if (r)
  837. return r;
  838. n->index = 0;
  839. c->depth++;
  840. if (c->prefetch_leaves || !leaf_node(c))
  841. prefetch_values(c);
  842. return 0;
  843. }
  844. static void pop_node(struct dm_btree_cursor *c)
  845. {
  846. c->depth--;
  847. unlock_block(c->info, c->nodes[c->depth].b);
  848. }
  849. static int inc_or_backtrack(struct dm_btree_cursor *c)
  850. {
  851. struct cursor_node *n;
  852. struct btree_node *bn;
  853. for (;;) {
  854. if (!c->depth)
  855. return -ENODATA;
  856. n = c->nodes + c->depth - 1;
  857. bn = dm_block_data(n->b);
  858. n->index++;
  859. if (n->index < le32_to_cpu(bn->header.nr_entries))
  860. break;
  861. pop_node(c);
  862. }
  863. return 0;
  864. }
  865. static int find_leaf(struct dm_btree_cursor *c)
  866. {
  867. int r = 0;
  868. struct cursor_node *n;
  869. struct btree_node *bn;
  870. __le64 value_le;
  871. for (;;) {
  872. n = c->nodes + c->depth - 1;
  873. bn = dm_block_data(n->b);
  874. if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
  875. break;
  876. memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
  877. r = push_node(c, le64_to_cpu(value_le));
  878. if (r) {
  879. DMERR("push_node failed");
  880. break;
  881. }
  882. }
  883. if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
  884. return -ENODATA;
  885. return r;
  886. }
  887. int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
  888. bool prefetch_leaves, struct dm_btree_cursor *c)
  889. {
  890. int r;
  891. c->info = info;
  892. c->root = root;
  893. c->depth = 0;
  894. c->prefetch_leaves = prefetch_leaves;
  895. r = push_node(c, root);
  896. if (r)
  897. return r;
  898. return find_leaf(c);
  899. }
  900. EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
  901. void dm_btree_cursor_end(struct dm_btree_cursor *c)
  902. {
  903. while (c->depth)
  904. pop_node(c);
  905. }
  906. EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
  907. int dm_btree_cursor_next(struct dm_btree_cursor *c)
  908. {
  909. int r = inc_or_backtrack(c);
  910. if (!r) {
  911. r = find_leaf(c);
  912. if (r)
  913. DMERR("find_leaf failed");
  914. }
  915. return r;
  916. }
  917. EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
  918. int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
  919. {
  920. if (c->depth) {
  921. struct cursor_node *n = c->nodes + c->depth - 1;
  922. struct btree_node *bn = dm_block_data(n->b);
  923. if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
  924. return -EINVAL;
  925. *key = le64_to_cpu(*key_ptr(bn, n->index));
  926. memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
  927. return 0;
  928. } else
  929. return -ENODATA;
  930. }
  931. EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);