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_api.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, minimbl;
205 struct public_ip_list *minip;
207 bool should_loop = true;
208 bool have_unassigned = true;
210 numnodes = ipalloc_state->num;
212 while (have_unassigned && should_loop) {
215 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
216 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES (UNASSIGNED)\n"));
222 /* loop over each unassigned ip. */
223 for (t = ipalloc_state->all_ips; t != NULL ; t = t->next) {
228 for (dstnode = 0; dstnode < numnodes; dstnode++) {
229 /* only check nodes that can actually takeover this ip */
230 if (!can_node_takeover_ip(ipalloc_state,
233 /* no it couldnt so skip to the next node */
237 dstdsum = ip_distance_2_sum(&(t->addr),
238 ipalloc_state->all_ips,
240 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
242 (" %s -> %d [+%d]\n",
243 ctdb_sock_addr_to_string(ipalloc_state,
246 dstimbl - lcp2_imbalances[dstnode]));
249 if ((minnode == -1) || (dstdsum < mindsum)) {
259 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
261 /* If we found one then assign it to the given node. */
263 minip->pnn = minnode;
264 lcp2_imbalances[minnode] = minimbl;
265 DEBUG(DEBUG_INFO,(" %s -> %d [+%d]\n",
266 ctdb_sock_addr_to_string(
273 /* There might be a better way but at least this is clear. */
274 have_unassigned = false;
275 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
277 have_unassigned = true;
282 /* We know if we have an unassigned addresses so we might as
285 if (have_unassigned) {
286 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
289 ("Failed to find node to cover ip %s\n",
290 ctdb_sock_addr_to_string(ipalloc_state,
297 /* LCP2 algorithm for rebalancing the cluster. Given a candidate node
298 * to move IPs from, determines the best IP/destination node
299 * combination to move from the source node.
301 static bool lcp2_failback_candidate(struct ipalloc_state *ipalloc_state,
303 uint32_t *lcp2_imbalances,
304 bool *rebalance_candidates)
306 int dstnode, mindstnode, numnodes;
307 uint32_t srcimbl, srcdsum, dstimbl, dstdsum;
308 uint32_t minsrcimbl, mindstimbl;
309 struct public_ip_list *minip;
310 struct public_ip_list *t;
312 /* Find an IP and destination node that best reduces imbalance. */
319 numnodes = ipalloc_state->num;
321 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
322 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES FROM %d [%d]\n",
323 srcnode, lcp2_imbalances[srcnode]));
325 for (t = ipalloc_state->all_ips; t != NULL; t = t->next) {
326 /* Only consider addresses on srcnode. */
327 if (t->pnn != srcnode) {
331 /* What is this IP address costing the source node? */
332 srcdsum = ip_distance_2_sum(&(t->addr),
333 ipalloc_state->all_ips,
335 srcimbl = lcp2_imbalances[srcnode] - srcdsum;
337 /* Consider this IP address would cost each potential
338 * destination node. Destination nodes are limited to
339 * those that are newly healthy, since we don't want
340 * to do gratuitous failover of IPs just to make minor
341 * balance improvements.
343 for (dstnode = 0; dstnode < numnodes; dstnode++) {
344 if (!rebalance_candidates[dstnode]) {
348 /* only check nodes that can actually takeover this ip */
349 if (!can_node_takeover_ip(ipalloc_state, dstnode,
351 /* no it couldnt so skip to the next node */
355 dstdsum = ip_distance_2_sum(&(t->addr),
356 ipalloc_state->all_ips,
358 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
359 DEBUG(DEBUG_DEBUG,(" %d [%d] -> %s -> %d [+%d]\n",
361 ctdb_sock_addr_to_string(
362 ipalloc_state, &(t->addr)),
365 if ((dstimbl < lcp2_imbalances[srcnode]) &&
366 (dstdsum < srcdsum) && \
367 ((mindstnode == -1) || \
368 ((srcimbl + dstimbl) < (minsrcimbl + mindstimbl)))) {
371 minsrcimbl = srcimbl;
372 mindstnode = dstnode;
373 mindstimbl = dstimbl;
377 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
379 if (mindstnode != -1) {
380 /* We found a move that makes things better... */
382 ("%d [%d] -> %s -> %d [+%d]\n",
383 srcnode, minsrcimbl - lcp2_imbalances[srcnode],
384 ctdb_sock_addr_to_string(ipalloc_state, &(minip->addr)),
385 mindstnode, mindstimbl - lcp2_imbalances[mindstnode]));
388 lcp2_imbalances[srcnode] = minsrcimbl;
389 lcp2_imbalances[mindstnode] = mindstimbl;
390 minip->pnn = mindstnode;
398 struct lcp2_imbalance_pnn {
403 static int lcp2_cmp_imbalance_pnn(const void * a, const void * b)
405 const struct lcp2_imbalance_pnn * lipa = (const struct lcp2_imbalance_pnn *) a;
406 const struct lcp2_imbalance_pnn * lipb = (const struct lcp2_imbalance_pnn *) b;
408 if (lipa->imbalance > lipb->imbalance) {
410 } else if (lipa->imbalance == lipb->imbalance) {
417 /* LCP2 algorithm for rebalancing the cluster. This finds the source
418 * node with the highest LCP2 imbalance, and then determines the best
419 * IP/destination node combination to move from the source node.
421 static void lcp2_failback(struct ipalloc_state *ipalloc_state,
422 uint32_t *lcp2_imbalances,
423 bool *rebalance_candidates)
426 struct lcp2_imbalance_pnn * lips;
429 numnodes = ipalloc_state->num;
432 /* Put the imbalances and nodes into an array, sort them and
433 * iterate through candidates. Usually the 1st one will be
434 * used, so this doesn't cost much...
436 DEBUG(DEBUG_DEBUG,("+++++++++++++++++++++++++++++++++++++++++\n"));
437 DEBUG(DEBUG_DEBUG,("Selecting most imbalanced node from:\n"));
438 lips = talloc_array(ipalloc_state, struct lcp2_imbalance_pnn, numnodes);
439 for (i = 0; i < numnodes; i++) {
440 lips[i].imbalance = lcp2_imbalances[i];
442 DEBUG(DEBUG_DEBUG,(" %d [%d]\n", i, lcp2_imbalances[i]));
444 qsort(lips, numnodes, sizeof(struct lcp2_imbalance_pnn),
445 lcp2_cmp_imbalance_pnn);
448 for (i = 0; i < numnodes; i++) {
449 /* This means that all nodes had 0 or 1 addresses, so
450 * can't be imbalanced.
452 if (lips[i].imbalance == 0) {
456 if (lcp2_failback_candidate(ipalloc_state,
459 rebalance_candidates)) {
471 bool ipalloc_lcp2(struct ipalloc_state *ipalloc_state)
473 uint32_t *lcp2_imbalances;
474 bool *rebalance_candidates;
475 int numnodes, num_rebalance_candidates, i;
478 unassign_unsuitable_ips(ipalloc_state);
480 if (!lcp2_init(ipalloc_state,
481 &lcp2_imbalances, &rebalance_candidates)) {
486 lcp2_allocate_unassigned(ipalloc_state, lcp2_imbalances);
488 /* If we don't want IPs to fail back then don't rebalance IPs. */
489 if (1 == ipalloc_state->no_ip_failback) {
493 /* It is only worth continuing if we have suitable target
494 * nodes to transfer IPs to. This check is much cheaper than
497 numnodes = ipalloc_state->num;
498 num_rebalance_candidates = 0;
499 for (i=0; i<numnodes; i++) {
500 if (rebalance_candidates[i]) {
501 num_rebalance_candidates++;
504 if (num_rebalance_candidates == 0) {
508 /* Now, try to make sure the ip adresses are evenly distributed
511 lcp2_failback(ipalloc_state, lcp2_imbalances, rebalance_candidates);