qnetd-algo-ffsplit.c 28 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886
  1. /*
  2. * Copyright (c) 2015-2019 Red Hat, Inc.
  3. *
  4. * All rights reserved.
  5. *
  6. * Author: Jan Friesse (jfriesse@redhat.com)
  7. *
  8. * This software licensed under BSD license, the text of which follows:
  9. *
  10. * Redistribution and use in source and binary forms, with or without
  11. * modification, are permitted provided that the following conditions are met:
  12. *
  13. * - Redistributions of source code must retain the above copyright notice,
  14. * this list of conditions and the following disclaimer.
  15. * - Redistributions in binary form must reproduce the above copyright notice,
  16. * this list of conditions and the following disclaimer in the documentation
  17. * and/or other materials provided with the distribution.
  18. * - Neither the name of the Red Hat, Inc. nor the names of its
  19. * contributors may be used to endorse or promote products derived from this
  20. * software without specific prior written permission.
  21. *
  22. * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  23. * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  24. * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  25. * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  26. * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  27. * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  28. * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
  29. * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
  30. * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  31. * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
  32. * THE POSSIBILITY OF SUCH DAMAGE.
  33. */
  34. #include <sys/types.h>
  35. #include <string.h>
  36. #include "log.h"
  37. #include "log-common.h"
  38. #include "qnetd-algo-ffsplit.h"
  39. #include "qnetd-log-debug.h"
  40. #include "qnetd-cluster-list.h"
  41. #include "qnetd-cluster.h"
  42. #include "qnetd-client-send.h"
  43. enum qnetd_algo_ffsplit_cluster_state {
  44. QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_CHANGE,
  45. QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_STABLE_MEMBERSHIP,
  46. QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_NACKS,
  47. QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_ACKS,
  48. };
  49. struct qnetd_algo_ffsplit_cluster_data {
  50. enum qnetd_algo_ffsplit_cluster_state cluster_state;
  51. const struct node_list *quorate_partition_node_list;
  52. };
  53. enum qnetd_algo_ffsplit_client_state {
  54. QNETD_ALGO_FFSPLIT_CLIENT_STATE_WAITING_FOR_CHANGE,
  55. QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK,
  56. QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK,
  57. };
  58. struct qnetd_algo_ffsplit_client_data {
  59. enum qnetd_algo_ffsplit_client_state client_state;
  60. uint32_t vote_info_expected_seq_num;
  61. };
  62. enum tlv_reply_error_code
  63. qnetd_algo_ffsplit_client_init(struct qnetd_client *client)
  64. {
  65. struct qnetd_algo_ffsplit_cluster_data *cluster_data;
  66. struct qnetd_algo_ffsplit_client_data *client_data;
  67. if (qnetd_cluster_size(client->cluster) == 1) {
  68. cluster_data = malloc(sizeof(*cluster_data));
  69. if (cluster_data == NULL) {
  70. log(LOG_ERR, "ffsplit: Can't initialize cluster data for client %s",
  71. client->addr_str);
  72. return (TLV_REPLY_ERROR_CODE_INTERNAL_ERROR);
  73. }
  74. memset(cluster_data, 0, sizeof(*cluster_data));
  75. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_CHANGE;
  76. cluster_data->quorate_partition_node_list = NULL;
  77. client->cluster->algorithm_data = cluster_data;
  78. }
  79. client_data = malloc(sizeof(*client_data));
  80. if (client_data == NULL) {
  81. log(LOG_ERR, "ffsplit: Can't initialize node data for client %s",
  82. client->addr_str);
  83. return (TLV_REPLY_ERROR_CODE_INTERNAL_ERROR);
  84. }
  85. memset(client_data, 0, sizeof(*client_data));
  86. client_data->client_state = QNETD_ALGO_FFSPLIT_CLIENT_STATE_WAITING_FOR_CHANGE;
  87. client->algorithm_data = client_data;
  88. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  89. }
  90. static int
  91. qnetd_algo_ffsplit_is_preferred_partition(const struct qnetd_client *client,
  92. const struct node_list *config_node_list, const struct node_list *membership_node_list)
  93. {
  94. uint32_t preferred_node_id;
  95. struct node_list_entry *node_entry;
  96. int case_processed;
  97. preferred_node_id = 0;
  98. case_processed = 0;
  99. switch (client->tie_breaker.mode) {
  100. case TLV_TIE_BREAKER_MODE_LOWEST:
  101. node_entry = TAILQ_FIRST(config_node_list);
  102. preferred_node_id = node_entry->node_id;
  103. TAILQ_FOREACH(node_entry, config_node_list, entries) {
  104. if (node_entry->node_id < preferred_node_id) {
  105. preferred_node_id = node_entry->node_id;
  106. }
  107. }
  108. case_processed = 1;
  109. break;
  110. case TLV_TIE_BREAKER_MODE_HIGHEST:
  111. node_entry = TAILQ_FIRST(config_node_list);
  112. preferred_node_id = node_entry->node_id;
  113. TAILQ_FOREACH(node_entry, config_node_list, entries) {
  114. if (node_entry->node_id > preferred_node_id) {
  115. preferred_node_id = node_entry->node_id;
  116. }
  117. }
  118. case_processed = 1;
  119. break;
  120. case TLV_TIE_BREAKER_MODE_NODE_ID:
  121. preferred_node_id = client->tie_breaker.node_id;
  122. case_processed = 1;
  123. break;
  124. }
  125. if (!case_processed) {
  126. log(LOG_CRIT, "qnetd_algo_ffsplit_is_preferred_partition unprocessed "
  127. "tie_breaker.mode");
  128. exit(1);
  129. }
  130. return (node_list_find_node_id(membership_node_list, preferred_node_id) != NULL);
  131. }
  132. static int
  133. qnetd_algo_ffsplit_is_membership_stable(const struct qnetd_client *client, int client_leaving,
  134. const struct tlv_ring_id *ring_id, const struct node_list *config_node_list,
  135. const struct node_list *membership_node_list)
  136. {
  137. const struct qnetd_client *iter_client1, *iter_client2;
  138. const struct node_list *config_node_list1, *config_node_list2;
  139. const struct node_list *membership_node_list1, *membership_node_list2;
  140. const struct node_list_entry *iter_node1, *iter_node2;
  141. const struct node_list_entry *iter_node3, *iter_node4;
  142. const struct tlv_ring_id *ring_id1, *ring_id2;
  143. /*
  144. * Test if all active clients share same config list.
  145. */
  146. TAILQ_FOREACH(iter_client1, &client->cluster->client_list, cluster_entries) {
  147. TAILQ_FOREACH(iter_client2, &client->cluster->client_list, cluster_entries) {
  148. if (iter_client1 == iter_client2) {
  149. continue;
  150. }
  151. if (iter_client1->node_id == client->node_id) {
  152. if (client_leaving) {
  153. continue;
  154. }
  155. config_node_list1 = config_node_list;
  156. } else {
  157. config_node_list1 = &iter_client1->configuration_node_list;
  158. }
  159. if (iter_client2->node_id == client->node_id) {
  160. if (client_leaving) {
  161. continue;
  162. }
  163. config_node_list2 = config_node_list;
  164. } else {
  165. config_node_list2 = &iter_client2->configuration_node_list;
  166. }
  167. /*
  168. * Walk thru all node ids in given config node list...
  169. */
  170. TAILQ_FOREACH(iter_node1, config_node_list1, entries) {
  171. /*
  172. * ... and try to find given node id in other list
  173. */
  174. iter_node2 = node_list_find_node_id(config_node_list2, iter_node1->node_id);
  175. if (iter_node2 == NULL) {
  176. /*
  177. * Node with iter_node1->node_id was not found in
  178. * config_node_list2 -> lists doesn't match
  179. */
  180. return (0);
  181. }
  182. }
  183. }
  184. }
  185. /*
  186. * Test if same partitions share same ring ids and membership node list
  187. */
  188. TAILQ_FOREACH(iter_client1, &client->cluster->client_list, cluster_entries) {
  189. if (iter_client1->node_id == client->node_id) {
  190. if (client_leaving) {
  191. continue;
  192. }
  193. membership_node_list1 = membership_node_list;
  194. ring_id1 = ring_id;
  195. } else {
  196. membership_node_list1 = &iter_client1->last_membership_node_list;
  197. ring_id1 = &iter_client1->last_ring_id;
  198. }
  199. /*
  200. * Walk thru all memberships nodes
  201. */
  202. TAILQ_FOREACH(iter_node1, membership_node_list1, entries) {
  203. /*
  204. * try to find client with given node id
  205. */
  206. iter_client2 = qnetd_cluster_find_client_by_node_id(client->cluster,
  207. iter_node1->node_id);
  208. if (iter_client2 == NULL) {
  209. /*
  210. * Client with given id is not connected
  211. */
  212. continue;
  213. }
  214. if (iter_client2->node_id == client->node_id) {
  215. if (client_leaving) {
  216. continue;
  217. }
  218. membership_node_list2 = membership_node_list;
  219. ring_id2 = ring_id;
  220. } else {
  221. membership_node_list2 = &iter_client2->last_membership_node_list;
  222. ring_id2 = &iter_client2->last_ring_id;
  223. }
  224. /*
  225. * Compare ring ids
  226. */
  227. if (!tlv_ring_id_eq(ring_id1, ring_id2)) {
  228. return (0);
  229. }
  230. /*
  231. * Now compare that membership node list equals, so walk thru all
  232. * members ...
  233. */
  234. TAILQ_FOREACH(iter_node3, membership_node_list1, entries) {
  235. /*
  236. * ... and try to find given node id in other membership node list
  237. */
  238. iter_node4 = node_list_find_node_id(membership_node_list2, iter_node3->node_id);
  239. if (iter_node4 == NULL) {
  240. /*
  241. * Node with iter_node3->node_id was not found in
  242. * membership_node_list2 -> lists doesn't match
  243. */
  244. return (0);
  245. }
  246. }
  247. }
  248. }
  249. return (1);
  250. }
  251. static void
  252. qnetd_algo_ffsplit_get_active_clients_in_partition_stats(const struct qnetd_client *client,
  253. const struct node_list *client_membership_node_list, enum tlv_heuristics client_heuristics,
  254. size_t *no_clients, size_t *no_heuristics_pass, size_t *no_heuristics_fail)
  255. {
  256. const struct node_list_entry *iter_node;
  257. const struct qnetd_client *iter_client;
  258. enum tlv_heuristics iter_heuristics;
  259. *no_clients = 0;
  260. *no_heuristics_pass = 0;
  261. *no_heuristics_fail = 0;
  262. if (client == NULL || client_membership_node_list == NULL) {
  263. return ;
  264. }
  265. TAILQ_FOREACH(iter_node, client_membership_node_list, entries) {
  266. iter_client = qnetd_cluster_find_client_by_node_id(client->cluster,
  267. iter_node->node_id);
  268. if (iter_client != NULL) {
  269. (*no_clients)++;
  270. if (iter_client == client) {
  271. iter_heuristics = client_heuristics;
  272. } else {
  273. iter_heuristics = iter_client->last_heuristics;
  274. }
  275. if (iter_heuristics == TLV_HEURISTICS_PASS) {
  276. (*no_heuristics_pass)++;
  277. } else if (iter_heuristics == TLV_HEURISTICS_FAIL) {
  278. (*no_heuristics_fail)++;
  279. }
  280. }
  281. }
  282. }
  283. /*
  284. * Compares two partitions. Return 1 if client1, config_node_list1, membership_node_list1 is
  285. * "better" than client2, config_node_list2, membership_node_list2
  286. */
  287. static int
  288. qnetd_algo_ffsplit_partition_cmp(const struct qnetd_client *client1,
  289. const struct node_list *config_node_list1, const struct node_list *membership_node_list1,
  290. enum tlv_heuristics heuristics_1,
  291. const struct qnetd_client *client2,
  292. const struct node_list *config_node_list2, const struct node_list *membership_node_list2,
  293. enum tlv_heuristics heuristics_2)
  294. {
  295. size_t part1_active_clients, part2_active_clients;
  296. size_t part1_no_heuristics_pass, part2_no_heuristics_pass;
  297. size_t part1_no_heuristics_fail, part2_no_heuristics_fail;
  298. size_t part1_score, part2_score;
  299. int res;
  300. res = -1;
  301. if (node_list_size(config_node_list1) % 2 != 0) {
  302. /*
  303. * Odd clusters never split into 50:50.
  304. */
  305. if (node_list_size(membership_node_list1) > node_list_size(config_node_list1) / 2) {
  306. res = 1; goto exit_res;
  307. } else {
  308. res = 0; goto exit_res;
  309. }
  310. } else {
  311. if (node_list_size(membership_node_list1) > node_list_size(config_node_list1) / 2) {
  312. res = 1; goto exit_res;
  313. } else if (node_list_size(membership_node_list1) < node_list_size(config_node_list1) / 2) {
  314. res = 0; goto exit_res;
  315. }
  316. /*
  317. * 50:50 split
  318. */
  319. /*
  320. * Check how many active clients are in partitions and heuristics results
  321. */
  322. qnetd_algo_ffsplit_get_active_clients_in_partition_stats(client1,
  323. membership_node_list1, heuristics_1, &part1_active_clients,
  324. &part1_no_heuristics_pass, &part1_no_heuristics_fail);
  325. qnetd_algo_ffsplit_get_active_clients_in_partition_stats(client2,
  326. membership_node_list2, heuristics_2, &part2_active_clients,
  327. &part2_no_heuristics_pass, &part2_no_heuristics_fail);
  328. /*
  329. * Partition can contain clients with one of 4 states:
  330. * 1. Not-connected to qnetd (D)
  331. * 2. Disabled heuristics (U)
  332. * 3. Enabled heuristics with pass result (P)
  333. * 4. Enabled heuristics with fail result (F)
  334. *
  335. * The question is, what partition should get vote is kind of hard with
  336. * so much states. Following simple "score" seems to be good enough, but may
  337. * be suboptimal in some cases. As and example let's say there are
  338. * 2 partitions with 4 nodes each. Partition 1 looks like PDDD and partition 2 looks
  339. * like FUUU. Partition 1 score is 1 + (1 - 0), partition 2 score is 4 + (0 - 1).
  340. * Partition 2 wins eventho there is one processor with failed heuristics.
  341. */
  342. part1_score = part1_active_clients + (part1_no_heuristics_pass - part1_no_heuristics_fail);
  343. part2_score = part2_active_clients + (part2_no_heuristics_pass - part2_no_heuristics_fail);
  344. if (part1_score > part2_score) {
  345. res = 1; goto exit_res;
  346. } else if (part1_score < part2_score) {
  347. res = 0; goto exit_res;
  348. }
  349. if (part1_active_clients > part2_active_clients) {
  350. res = 1; goto exit_res;
  351. } else if (part1_active_clients < part2_active_clients) {
  352. res = 0; goto exit_res;
  353. }
  354. /*
  355. * Number of active clients in both partitions equals. Use tie-breaker.
  356. */
  357. if (qnetd_algo_ffsplit_is_preferred_partition(client1, config_node_list1,
  358. membership_node_list1)) {
  359. res = 1; goto exit_res;
  360. } else {
  361. res = 0; goto exit_res;
  362. }
  363. }
  364. exit_res:
  365. if (res == -1) {
  366. log(LOG_CRIT, "qnetd_algo_ffsplit_partition_cmp unhandled case");
  367. exit(1);
  368. /* NOTREACHED */
  369. }
  370. return (res);
  371. }
  372. /*
  373. * Select best partition for given client->cluster.
  374. * If there is no partition which could become quorate, NULL is returned
  375. */
  376. static const struct node_list *
  377. qnetd_algo_ffsplit_select_partition(const struct qnetd_client *client, int client_leaving,
  378. const struct node_list *config_node_list, const struct node_list *membership_node_list,
  379. enum tlv_heuristics client_heuristics)
  380. {
  381. const struct qnetd_client *iter_client;
  382. const struct qnetd_client *best_client;
  383. const struct node_list *best_config_node_list, *best_membership_node_list;
  384. const struct node_list *iter_config_node_list, *iter_membership_node_list;
  385. enum tlv_heuristics iter_heuristics, best_heuristics;
  386. best_client = NULL;
  387. best_config_node_list = best_membership_node_list = NULL;
  388. best_heuristics = TLV_HEURISTICS_UNDEFINED;
  389. /*
  390. * Get highest score
  391. */
  392. TAILQ_FOREACH(iter_client, &client->cluster->client_list, cluster_entries) {
  393. if (iter_client->node_id == client->node_id) {
  394. if (client_leaving) {
  395. continue;
  396. }
  397. iter_config_node_list = config_node_list;
  398. iter_membership_node_list = membership_node_list;
  399. iter_heuristics = client_heuristics;
  400. } else {
  401. iter_config_node_list = &iter_client->configuration_node_list;
  402. iter_membership_node_list = &iter_client->last_membership_node_list;
  403. iter_heuristics = iter_client->last_heuristics;
  404. }
  405. if (qnetd_algo_ffsplit_partition_cmp(iter_client, iter_config_node_list,
  406. iter_membership_node_list, iter_heuristics, best_client, best_config_node_list,
  407. best_membership_node_list, best_heuristics) > 0) {
  408. best_client = iter_client;
  409. best_config_node_list = iter_config_node_list;
  410. best_membership_node_list = iter_membership_node_list;
  411. best_heuristics = iter_heuristics;
  412. }
  413. }
  414. return (best_membership_node_list);
  415. }
  416. /*
  417. * Update state of all nodes to match quorate_partition_node_list
  418. */
  419. static void
  420. qnetd_algo_ffsplit_update_nodes_state(struct qnetd_client *client, int client_leaving,
  421. const struct node_list *quorate_partition_node_list)
  422. {
  423. const struct qnetd_client *iter_client;
  424. struct qnetd_algo_ffsplit_client_data *iter_client_data;
  425. TAILQ_FOREACH(iter_client, &client->cluster->client_list, cluster_entries) {
  426. iter_client_data = (struct qnetd_algo_ffsplit_client_data *)iter_client->algorithm_data;
  427. if (iter_client->node_id == client->node_id && client_leaving) {
  428. iter_client_data->client_state = QNETD_ALGO_FFSPLIT_CLIENT_STATE_WAITING_FOR_CHANGE;
  429. continue;
  430. }
  431. if (quorate_partition_node_list == NULL ||
  432. node_list_find_node_id(quorate_partition_node_list, iter_client->node_id) == NULL) {
  433. iter_client_data->client_state = QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK;
  434. } else {
  435. iter_client_data->client_state = QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK;
  436. }
  437. }
  438. }
  439. /*
  440. * Send vote info. If client_leaving is set, client is ignored. if send_acks
  441. * is set, only ACK votes are sent (nodes in QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK state),
  442. * otherwise only NACK votes are sent (nodes in QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK state)
  443. *
  444. * Returns number of send votes
  445. */
  446. static size_t
  447. qnetd_algo_ffsplit_send_votes(struct qnetd_client *client, int client_leaving,
  448. const struct tlv_ring_id *ring_id, int send_acks)
  449. {
  450. size_t sent_votes;
  451. struct qnetd_client *iter_client;
  452. struct qnetd_algo_ffsplit_client_data *iter_client_data;
  453. const struct tlv_ring_id *ring_id_to_send;
  454. enum tlv_vote vote_to_send;
  455. sent_votes = 0;
  456. TAILQ_FOREACH(iter_client, &client->cluster->client_list, cluster_entries) {
  457. if (iter_client->node_id == client->node_id) {
  458. if (client_leaving) {
  459. continue;
  460. }
  461. ring_id_to_send = ring_id;
  462. } else {
  463. ring_id_to_send = &iter_client->last_ring_id;
  464. }
  465. iter_client_data = (struct qnetd_algo_ffsplit_client_data *)iter_client->algorithm_data;
  466. vote_to_send = TLV_VOTE_UNDEFINED;
  467. if (send_acks) {
  468. if (iter_client_data->client_state == QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK) {
  469. vote_to_send = TLV_VOTE_ACK;
  470. }
  471. } else {
  472. if (iter_client_data->client_state == QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK) {
  473. vote_to_send = TLV_VOTE_NACK;
  474. }
  475. }
  476. if (vote_to_send != TLV_VOTE_UNDEFINED) {
  477. iter_client_data->vote_info_expected_seq_num++;
  478. sent_votes++;
  479. if (qnetd_client_send_vote_info(iter_client,
  480. iter_client_data->vote_info_expected_seq_num, ring_id_to_send,
  481. vote_to_send) == -1) {
  482. client->schedule_disconnect = 1;
  483. }
  484. }
  485. }
  486. return (sent_votes);
  487. }
  488. /*
  489. * Return number of clients in QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK state if sending_acks is
  490. * set or number of nodes in QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK state if sending_acks is
  491. * not set
  492. */
  493. static size_t
  494. qnetd_algo_ffsplit_no_clients_in_sending_state(struct qnetd_client *client, int sending_acks)
  495. {
  496. size_t no_clients;
  497. struct qnetd_client *iter_client;
  498. struct qnetd_algo_ffsplit_client_data *iter_client_data;
  499. no_clients = 0;
  500. TAILQ_FOREACH(iter_client, &client->cluster->client_list, cluster_entries) {
  501. iter_client_data = (struct qnetd_algo_ffsplit_client_data *)iter_client->algorithm_data;
  502. if (sending_acks &&
  503. iter_client_data->client_state == QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_ACK) {
  504. no_clients++;
  505. }
  506. if (!sending_acks &&
  507. iter_client_data->client_state == QNETD_ALGO_FFSPLIT_CLIENT_STATE_SENDING_NACK) {
  508. no_clients++;
  509. }
  510. }
  511. return (no_clients);
  512. }
  513. static enum tlv_vote
  514. qnetd_algo_ffsplit_do(struct qnetd_client *client, int client_leaving,
  515. const struct tlv_ring_id *ring_id, const struct node_list *config_node_list,
  516. const struct node_list *membership_node_list, enum tlv_heuristics client_heuristics)
  517. {
  518. struct qnetd_algo_ffsplit_cluster_data *cluster_data;
  519. const struct node_list *quorate_partition_node_list;
  520. cluster_data = (struct qnetd_algo_ffsplit_cluster_data *)client->cluster->algorithm_data;
  521. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_STABLE_MEMBERSHIP;
  522. if (!qnetd_algo_ffsplit_is_membership_stable(client, client_leaving,
  523. ring_id, config_node_list, membership_node_list)) {
  524. /*
  525. * Wait until membership is stable
  526. */
  527. log(LOG_DEBUG, "ffsplit: Membership for cluster %s is not yet stable", client->cluster_name);
  528. return (TLV_VOTE_WAIT_FOR_REPLY);
  529. }
  530. log(LOG_DEBUG, "ffsplit: Membership for cluster %s is now stable", client->cluster_name);
  531. quorate_partition_node_list = qnetd_algo_ffsplit_select_partition(client, client_leaving,
  532. config_node_list, membership_node_list, client_heuristics);
  533. cluster_data->quorate_partition_node_list = quorate_partition_node_list;
  534. if (quorate_partition_node_list == NULL) {
  535. log(LOG_DEBUG, "ffsplit: No quorate partition was selected");
  536. } else {
  537. log(LOG_DEBUG, "ffsplit: Quorate partition selected");
  538. log_common_debug_dump_node_list(quorate_partition_node_list);
  539. }
  540. qnetd_algo_ffsplit_update_nodes_state(client, client_leaving, quorate_partition_node_list);
  541. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_NACKS;
  542. if (qnetd_algo_ffsplit_send_votes(client, client_leaving, ring_id, 0) == 0) {
  543. log(LOG_DEBUG, "ffsplit: No client gets NACK");
  544. /*
  545. * No one gets nack -> send acks
  546. */
  547. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_ACKS;
  548. if (qnetd_algo_ffsplit_send_votes(client, client_leaving, ring_id, 1) == 0) {
  549. log(LOG_DEBUG, "ffsplit: No client gets ACK");
  550. /*
  551. * No one gets acks -> finished
  552. */
  553. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_CHANGE;
  554. }
  555. }
  556. return (TLV_VOTE_NO_CHANGE);
  557. }
  558. enum tlv_reply_error_code
  559. qnetd_algo_ffsplit_config_node_list_received(struct qnetd_client *client,
  560. uint32_t msg_seq_num, int config_version_set, uint64_t config_version,
  561. const struct node_list *nodes, int initial, enum tlv_vote *result_vote)
  562. {
  563. if (node_list_size(nodes) == 0) {
  564. /*
  565. * Empty node list shouldn't happen
  566. */
  567. log(LOG_ERR, "ffsplit: Received empty config node list for client %s",
  568. client->addr_str);
  569. return (TLV_REPLY_ERROR_CODE_INVALID_CONFIG_NODE_LIST);
  570. }
  571. if (node_list_find_node_id(nodes, client->node_id) == NULL) {
  572. /*
  573. * Current node is not in node list
  574. */
  575. log(LOG_ERR, "ffsplit: Received config node list without client %s",
  576. client->addr_str);
  577. return (TLV_REPLY_ERROR_CODE_INVALID_CONFIG_NODE_LIST);
  578. }
  579. if (initial || node_list_size(&client->last_membership_node_list) == 0) {
  580. /*
  581. * Initial node list -> membership is going to be send by client
  582. */
  583. *result_vote = TLV_VOTE_ASK_LATER;
  584. } else {
  585. *result_vote = qnetd_algo_ffsplit_do(client, 0, &client->last_ring_id,
  586. nodes, &client->last_membership_node_list, client->last_heuristics);
  587. }
  588. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  589. }
  590. /*
  591. * Called after client sent membership node list.
  592. * All client fields are already set. Nodes is actual node list.
  593. * msg_seq_num is 32-bit number set by client. If client sent config file version,
  594. * config_version_set is set to 1 and config_version contains valid config file version.
  595. * ring_id and quorate are copied from client votequorum callback.
  596. *
  597. * Function has to return result_vote. This can be one of ack/nack, ask_later (client
  598. * should ask later for a vote) or wait_for_reply (client should wait for reply).
  599. *
  600. * Return TLV_REPLY_ERROR_CODE_NO_ERROR on success, different TLV_REPLY_ERROR_CODE_*
  601. * on failure (error is send back to client)
  602. */
  603. enum tlv_reply_error_code
  604. qnetd_algo_ffsplit_membership_node_list_received(struct qnetd_client *client,
  605. uint32_t msg_seq_num, const struct tlv_ring_id *ring_id,
  606. const struct node_list *nodes, enum tlv_heuristics heuristics, enum tlv_vote *result_vote)
  607. {
  608. if (node_list_size(nodes) == 0) {
  609. /*
  610. * Empty node list shouldn't happen
  611. */
  612. log(LOG_ERR, "ffsplit: Received empty membership node list for client %s",
  613. client->addr_str);
  614. return (TLV_REPLY_ERROR_CODE_INVALID_MEMBERSHIP_NODE_LIST);
  615. }
  616. if (node_list_find_node_id(nodes, client->node_id) == NULL) {
  617. /*
  618. * Current node is not in node list
  619. */
  620. log(LOG_ERR, "ffsplit: Received membership node list without client %s",
  621. client->addr_str);
  622. return (TLV_REPLY_ERROR_CODE_INVALID_MEMBERSHIP_NODE_LIST);
  623. }
  624. if (node_list_size(&client->configuration_node_list) == 0) {
  625. /*
  626. * Config node list not received -> it's going to be sent later
  627. */
  628. *result_vote = TLV_VOTE_ASK_LATER;
  629. } else {
  630. *result_vote = qnetd_algo_ffsplit_do(client, 0, ring_id,
  631. &client->configuration_node_list, nodes, heuristics);
  632. }
  633. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  634. }
  635. enum tlv_reply_error_code
  636. qnetd_algo_ffsplit_quorum_node_list_received(struct qnetd_client *client,
  637. uint32_t msg_seq_num, enum tlv_quorate quorate, const struct node_list *nodes,
  638. enum tlv_vote *result_vote)
  639. {
  640. /*
  641. * Quorum node list is informative -> no change
  642. */
  643. *result_vote = TLV_VOTE_NO_CHANGE;
  644. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  645. }
  646. void
  647. qnetd_algo_ffsplit_client_disconnect(struct qnetd_client *client, int server_going_down)
  648. {
  649. (void)qnetd_algo_ffsplit_do(client, 1, &client->last_ring_id,
  650. &client->configuration_node_list, &client->last_membership_node_list,
  651. client->last_heuristics);
  652. free(client->algorithm_data);
  653. if (qnetd_cluster_size(client->cluster) == 1) {
  654. /*
  655. * Last client in the cluster
  656. */
  657. free(client->cluster->algorithm_data);
  658. }
  659. }
  660. enum tlv_reply_error_code
  661. qnetd_algo_ffsplit_ask_for_vote_received(struct qnetd_client *client, uint32_t msg_seq_num,
  662. enum tlv_vote *result_vote)
  663. {
  664. /*
  665. * Ask for vote is not supported in current algorithm
  666. */
  667. return (TLV_REPLY_ERROR_CODE_UNSUPPORTED_DECISION_ALGORITHM_MESSAGE);
  668. }
  669. enum tlv_reply_error_code
  670. qnetd_algo_ffsplit_vote_info_reply_received(struct qnetd_client *client, uint32_t msg_seq_num)
  671. {
  672. struct qnetd_algo_ffsplit_cluster_data *cluster_data;
  673. struct qnetd_algo_ffsplit_client_data *client_data;
  674. cluster_data = (struct qnetd_algo_ffsplit_cluster_data *)client->cluster->algorithm_data;
  675. client_data = (struct qnetd_algo_ffsplit_client_data *)client->algorithm_data;
  676. if (client_data->vote_info_expected_seq_num != msg_seq_num) {
  677. log(LOG_DEBUG, "ffsplit: Received old vote info reply from client %s",
  678. client->addr_str);
  679. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  680. }
  681. client_data->client_state = QNETD_ALGO_FFSPLIT_CLIENT_STATE_WAITING_FOR_CHANGE;
  682. if (cluster_data->cluster_state != QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_NACKS &&
  683. cluster_data->cluster_state != QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_ACKS) {
  684. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  685. }
  686. if (cluster_data->cluster_state == QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_NACKS) {
  687. if (qnetd_algo_ffsplit_no_clients_in_sending_state(client, 0) == 0) {
  688. log(LOG_DEBUG, "ffsplit: All NACK votes sent for cluster %s",
  689. client->cluster_name);
  690. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_SENDING_ACKS;
  691. if (qnetd_algo_ffsplit_send_votes(client, 0, &client->last_ring_id, 1) == 0) {
  692. log(LOG_DEBUG, "ffsplit: No client gets ACK");
  693. /*
  694. * No one gets acks -> finished
  695. */
  696. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_CHANGE;
  697. }
  698. }
  699. } else {
  700. if (qnetd_algo_ffsplit_no_clients_in_sending_state(client, 1) == 0) {
  701. log(LOG_DEBUG, "ffsplit: All ACK votes sent for cluster %s",
  702. client->cluster_name);
  703. cluster_data->cluster_state = QNETD_ALGO_FFSPLIT_CLUSTER_STATE_WAITING_FOR_CHANGE;
  704. }
  705. }
  706. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  707. }
  708. enum tlv_reply_error_code
  709. qnetd_algo_ffsplit_heuristics_change_received(struct qnetd_client *client, uint32_t msg_seq_num,
  710. enum tlv_heuristics heuristics, enum tlv_vote *result_vote)
  711. {
  712. if (node_list_size(&client->configuration_node_list) == 0 ||
  713. node_list_size(&client->last_membership_node_list) == 0) {
  714. /*
  715. * Config or membership node list not received -> it's going to be sent later
  716. */
  717. *result_vote = TLV_VOTE_ASK_LATER;
  718. } else {
  719. *result_vote = qnetd_algo_ffsplit_do(client, 0, &client->last_ring_id,
  720. &client->configuration_node_list, &client->last_membership_node_list,
  721. heuristics);
  722. }
  723. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  724. }
  725. enum tlv_reply_error_code
  726. qnetd_algo_ffsplit_timer_callback(struct qnetd_client *client, int *reschedule_timer,
  727. int *send_vote, enum tlv_vote *result_vote)
  728. {
  729. return (TLV_REPLY_ERROR_CODE_NO_ERROR);
  730. }
  731. static struct qnetd_algorithm qnetd_algo_ffsplit = {
  732. .init = qnetd_algo_ffsplit_client_init,
  733. .config_node_list_received = qnetd_algo_ffsplit_config_node_list_received,
  734. .membership_node_list_received = qnetd_algo_ffsplit_membership_node_list_received,
  735. .quorum_node_list_received = qnetd_algo_ffsplit_quorum_node_list_received,
  736. .client_disconnect = qnetd_algo_ffsplit_client_disconnect,
  737. .ask_for_vote_received = qnetd_algo_ffsplit_ask_for_vote_received,
  738. .vote_info_reply_received = qnetd_algo_ffsplit_vote_info_reply_received,
  739. .heuristics_change_received = qnetd_algo_ffsplit_heuristics_change_received,
  740. .timer_callback = qnetd_algo_ffsplit_timer_callback,
  741. };
  742. enum tlv_reply_error_code qnetd_algo_ffsplit_register()
  743. {
  744. return (qnetd_algorithm_register(TLV_DECISION_ALGORITHM_TYPE_FFSPLIT, &qnetd_algo_ffsplit));
  745. }