4 Copyright (C) Ronnie Sahlberg 2007
5 Copyright (C) Andrew Tridgell 2007
6 Copyright (C) Martin Schwenke 2011
8 This program is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3 of the License, or
11 (at your option) any later version.
13 This program is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with this program; if not, see <http://www.gnu.org/licenses/>.
23 #include "system/network.h"
25 #include "lib/util/debug.h"
26 #include "common/logging.h"
28 #include "protocol/protocol_util.h"
30 #include "server/ipalloc_private.h"
33 * This is the length of the longtest common prefix between the IPs.
34 * It is calculated by XOR-ing the 2 IPs together and counting the
35 * number of leading zeroes. The implementation means that all
36 * addresses end up being 128 bits long.
38 * FIXME? Should we consider IPv4 and IPv6 separately given that the
39 * 12 bytes of 0 prefix padding will hurt the algorithm if there are
40 * lots of nodes and IP addresses?
42 static uint32_t ip_distance(ctdb_sock_addr *ip1, ctdb_sock_addr *ip2)
44 uint32_t ip1_k[IP_KEYLEN];
49 uint32_t distance = 0;
51 memcpy(ip1_k, ip_key(ip1), sizeof(ip1_k));
53 for (i=0; i<IP_KEYLEN; i++) {
58 /* Count number of leading zeroes.
59 * FIXME? This could be optimised...
61 while ((x & (1 << 31)) == 0) {
71 /* Calculate the IP distance for the given IP relative to IPs on the
72 given node. The ips argument is generally the all_ips variable
73 used in the main part of the algorithm.
75 static uint32_t ip_distance_2_sum(ctdb_sock_addr *ip,
76 struct public_ip_list *ips,
79 struct public_ip_list *t;
84 for (t = ips; t != NULL; t = t->next) {
89 /* Optimisation: We never calculate the distance
90 * between an address and itself. This allows us to
91 * calculate the effect of removing an address from a
92 * node by simply calculating the distance between
93 * that address and all of the exitsing addresses.
94 * Moreover, we assume that we're only ever dealing
95 * with addresses from all_ips so we can identify an
96 * address via a pointer rather than doing a more
97 * expensive address comparison. */
98 if (&(t->addr) == ip) {
102 d = ip_distance(ip, &(t->addr));
103 sum += d * d; /* Cheaper than pulling in math.h :-) */
109 /* Return the LCP2 imbalance metric for addresses currently assigned
112 static uint32_t lcp2_imbalance(struct public_ip_list * all_ips, int pnn)
114 struct public_ip_list *t;
116 uint32_t imbalance = 0;
118 for (t = all_ips; t != NULL; t = t->next) {
122 /* Pass the rest of the IPs rather than the whole
125 imbalance += ip_distance_2_sum(&(t->addr), t->next, pnn);
131 static bool lcp2_init(struct ipalloc_state *ipalloc_state,
132 uint32_t **lcp2_imbalances,
133 bool **rebalance_candidates)
136 struct public_ip_list *t;
138 numnodes = ipalloc_state->num;
140 *rebalance_candidates = talloc_array(ipalloc_state, bool, numnodes);
141 if (*rebalance_candidates == NULL) {
142 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
145 *lcp2_imbalances = talloc_array(ipalloc_state, uint32_t, numnodes);
146 if (*lcp2_imbalances == NULL) {
147 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
151 for (i=0; i<numnodes; i++) {
152 (*lcp2_imbalances)[i] =
153 lcp2_imbalance(ipalloc_state->all_ips, i);
154 /* First step: assume all nodes are candidates */
155 (*rebalance_candidates)[i] = true;
158 /* 2nd step: if a node has IPs assigned then it must have been
159 * healthy before, so we remove it from consideration. This
160 * is overkill but is all we have because we don't maintain
161 * state between takeover runs. An alternative would be to
162 * keep state and invalidate it every time the recovery master
165 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
167 (*rebalance_candidates)[t->pnn] = false;
171 /* 3rd step: if a node is forced to re-balance then
172 we allow failback onto the node */
173 if (ipalloc_state->force_rebalance_nodes == NULL) {
177 i < talloc_array_length(ipalloc_state->force_rebalance_nodes);
179 uint32_t pnn = ipalloc_state->force_rebalance_nodes[i];
180 if (pnn >= numnodes) {
182 (__location__ "unknown node %u\n", pnn));
187 ("Forcing rebalancing of IPs to node %u\n", pnn));
188 (*rebalance_candidates)[pnn] = true;
194 /* Allocate any unassigned addresses using the LCP2 algorithm to find
195 * the IP/node combination that will cost the least.
197 static void lcp2_allocate_unassigned(struct ipalloc_state *ipalloc_state,
198 uint32_t *lcp2_imbalances)
200 struct public_ip_list *t;
201 int dstnode, numnodes;
204 uint32_t mindsum, dstdsum, dstimbl;
205 uint32_t minimbl = 0;
206 struct public_ip_list *minip;
208 bool should_loop = true;
209 bool have_unassigned = true;
211 numnodes = ipalloc_state->num;
213 while (have_unassigned && should_loop) {
216 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
217 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES (UNASSIGNED)\n"));
223 /* loop over each unassigned ip. */
224 for (t = ipalloc_state->all_ips; t != NULL ; t = t->next) {
229 for (dstnode = 0; dstnode < numnodes; dstnode++) {
230 /* only check nodes that can actually takeover this ip */
231 if (!can_node_takeover_ip(ipalloc_state,
234 /* no it couldnt so skip to the next node */
238 dstdsum = ip_distance_2_sum(&(t->addr),
239 ipalloc_state->all_ips,
241 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
243 (" %s -> %d [+%d]\n",
244 ctdb_sock_addr_to_string(ipalloc_state,
248 dstimbl - lcp2_imbalances[dstnode]));
251 if ((minnode == -1) || (dstdsum < mindsum)) {
261 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
263 /* If we found one then assign it to the given node. */
265 minip->pnn = minnode;
266 lcp2_imbalances[minnode] = minimbl;
267 DEBUG(DEBUG_INFO,(" %s -> %d [+%d]\n",
268 ctdb_sock_addr_to_string(
270 &(minip->addr), false),
275 /* There might be a better way but at least this is clear. */
276 have_unassigned = false;
277 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
279 have_unassigned = true;
284 /* We know if we have an unassigned addresses so we might as
287 if (have_unassigned) {
288 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
291 ("Failed to find node to cover ip %s\n",
292 ctdb_sock_addr_to_string(ipalloc_state,
300 /* LCP2 algorithm for rebalancing the cluster. Given a candidate node
301 * to move IPs from, determines the best IP/destination node
302 * combination to move from the source node.
304 static bool lcp2_failback_candidate(struct ipalloc_state *ipalloc_state,
306 uint32_t *lcp2_imbalances,
307 bool *rebalance_candidates)
309 int dstnode, mindstnode, numnodes;
310 uint32_t srcimbl, srcdsum, dstimbl, dstdsum;
311 uint32_t minsrcimbl, mindstimbl;
312 struct public_ip_list *minip;
313 struct public_ip_list *t;
315 /* Find an IP and destination node that best reduces imbalance. */
322 numnodes = ipalloc_state->num;
324 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
325 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES FROM %d [%d]\n",
326 srcnode, lcp2_imbalances[srcnode]));
328 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
329 /* Only consider addresses on srcnode. */
330 if (t->pnn != srcnode) {
334 /* What is this IP address costing the source node? */
335 srcdsum = ip_distance_2_sum(&(t->addr),
336 ipalloc_state->all_ips,
338 srcimbl = lcp2_imbalances[srcnode] - srcdsum;
340 /* Consider this IP address would cost each potential
341 * destination node. Destination nodes are limited to
342 * those that are newly healthy, since we don't want
343 * to do gratuitous failover of IPs just to make minor
344 * balance improvements.
346 for (dstnode = 0; dstnode < numnodes; dstnode++) {
347 if (!rebalance_candidates[dstnode]) {
351 /* only check nodes that can actually takeover this ip */
352 if (!can_node_takeover_ip(ipalloc_state, dstnode,
354 /* no it couldnt so skip to the next node */
358 dstdsum = ip_distance_2_sum(&(t->addr),
359 ipalloc_state->all_ips,
361 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
362 DEBUG(DEBUG_DEBUG,(" %d [%d] -> %s -> %d [+%d]\n",
364 ctdb_sock_addr_to_string(
369 if ((dstimbl < lcp2_imbalances[srcnode]) &&
370 (dstdsum < srcdsum) && \
371 ((mindstnode == -1) || \
372 ((srcimbl + dstimbl) < (minsrcimbl + mindstimbl)))) {
375 minsrcimbl = srcimbl;
376 mindstnode = dstnode;
377 mindstimbl = dstimbl;
381 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
383 if (mindstnode != -1) {
384 /* We found a move that makes things better... */
386 ("%d [%d] -> %s -> %d [+%d]\n",
387 srcnode, minsrcimbl - lcp2_imbalances[srcnode],
388 ctdb_sock_addr_to_string(ipalloc_state,
389 &(minip->addr), false),
390 mindstnode, mindstimbl - lcp2_imbalances[mindstnode]));
393 lcp2_imbalances[srcnode] = minsrcimbl;
394 lcp2_imbalances[mindstnode] = mindstimbl;
395 minip->pnn = mindstnode;
403 struct lcp2_imbalance_pnn {
408 static int lcp2_cmp_imbalance_pnn(const void * a, const void * b)
410 const struct lcp2_imbalance_pnn * lipa = (const struct lcp2_imbalance_pnn *) a;
411 const struct lcp2_imbalance_pnn * lipb = (const struct lcp2_imbalance_pnn *) b;
413 if (lipa->imbalance > lipb->imbalance) {
415 } else if (lipa->imbalance == lipb->imbalance) {
422 /* LCP2 algorithm for rebalancing the cluster. This finds the source
423 * node with the highest LCP2 imbalance, and then determines the best
424 * IP/destination node combination to move from the source node.
426 static void lcp2_failback(struct ipalloc_state *ipalloc_state,
427 uint32_t *lcp2_imbalances,
428 bool *rebalance_candidates)
431 struct lcp2_imbalance_pnn * lips;
434 numnodes = ipalloc_state->num;
437 /* Put the imbalances and nodes into an array, sort them and
438 * iterate through candidates. Usually the 1st one will be
439 * used, so this doesn't cost much...
441 DEBUG(DEBUG_DEBUG,("+++++++++++++++++++++++++++++++++++++++++\n"));
442 DEBUG(DEBUG_DEBUG,("Selecting most imbalanced node from:\n"));
443 lips = talloc_array(ipalloc_state, struct lcp2_imbalance_pnn, numnodes);
444 for (i = 0; i < numnodes; i++) {
445 lips[i].imbalance = lcp2_imbalances[i];
447 DEBUG(DEBUG_DEBUG,(" %d [%d]\n", i, lcp2_imbalances[i]));
449 qsort(lips, numnodes, sizeof(struct lcp2_imbalance_pnn),
450 lcp2_cmp_imbalance_pnn);
453 for (i = 0; i < numnodes; i++) {
454 /* This means that all nodes had 0 or 1 addresses, so
455 * can't be imbalanced.
457 if (lips[i].imbalance == 0) {
461 if (lcp2_failback_candidate(ipalloc_state,
464 rebalance_candidates)) {
476 bool ipalloc_lcp2(struct ipalloc_state *ipalloc_state)
478 uint32_t *lcp2_imbalances;
479 bool *rebalance_candidates;
480 int numnodes, num_rebalance_candidates, i;
483 unassign_unsuitable_ips(ipalloc_state);
485 if (!lcp2_init(ipalloc_state,
486 &lcp2_imbalances, &rebalance_candidates)) {
491 lcp2_allocate_unassigned(ipalloc_state, lcp2_imbalances);
493 /* If we don't want IPs to fail back then don't rebalance IPs. */
494 if (ipalloc_state->no_ip_failback) {
498 /* It is only worth continuing if we have suitable target
499 * nodes to transfer IPs to. This check is much cheaper than
502 numnodes = ipalloc_state->num;
503 num_rebalance_candidates = 0;
504 for (i=0; i<numnodes; i++) {
505 if (rebalance_candidates[i]) {
506 num_rebalance_candidates++;
509 if (num_rebalance_candidates == 0) {
513 /* Now, try to make sure the ip adresses are evenly distributed
516 lcp2_failback(ipalloc_state, lcp2_imbalances, rebalance_candidates);