ctdb/server/ipalloc.c

   1 /*
   2    ctdb ip takeover code
   3
   4    Copyright (C) Ronnie Sahlberg  2007
   5    Copyright (C) Andrew Tridgell  2007
   6    Copyright (C) Martin Schwenke  2011
   7
   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.
  12
  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.
  17
  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/>.
  20 */
  21
  22 #include "replace.h"
  23 #include "system/network.h"
  24
  25 #include <talloc.h>
  26
  27 #include "lib/util/debug.h"
  28
  29 #include "common/logging.h"
  30 #include "common/rb_tree.h"
  31
  32 #include "protocol/protocol_util.h"
  33
  34 #include "server/ipalloc_private.h"
  35
  36 /* Initialise main ipalloc state and sub-structures */
  37 struct ipalloc_state *
  38 ipalloc_state_init(TALLOC_CTX *mem_ctx,
  39                    uint32_t num_nodes,
  40                    enum ipalloc_algorithm algorithm,
  41                    bool no_ip_takeover,
  42                    bool no_ip_failback,
  43                    uint32_t *force_rebalance_nodes)
  44 {
  45         struct ipalloc_state *ipalloc_state =
  46                 talloc_zero(mem_ctx, struct ipalloc_state);
  47         if (ipalloc_state == NULL) {
  48                 DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
  49                 return NULL;
  50         }
  51
  52         ipalloc_state->num = num_nodes;
  53
  54         ipalloc_state->algorithm = algorithm;
  55         ipalloc_state->no_ip_takeover = no_ip_takeover;
  56         ipalloc_state->no_ip_failback = no_ip_failback;
  57         ipalloc_state->force_rebalance_nodes = force_rebalance_nodes;
  58
  59         return ipalloc_state;
  60 }
  61
  62 static void *add_ip_callback(void *parm, void *data)
  63 {
  64         struct public_ip_list *this_ip = parm;
  65         struct public_ip_list *prev_ip = data;
  66
  67         if (prev_ip == NULL) {
  68                 return parm;
  69         }
  70         if (this_ip->pnn == CTDB_UNKNOWN_PNN) {
  71                 this_ip->pnn = prev_ip->pnn;
  72         }
  73
  74         return parm;
  75 }
  76
  77 static int getips_count_callback(void *param, void *data)
  78 {
  79         struct public_ip_list **ip_list = (struct public_ip_list **)param;
  80         struct public_ip_list *new_ip = (struct public_ip_list *)data;
  81
  82         new_ip->next = *ip_list;
  83         *ip_list     = new_ip;
  84         return 0;
  85 }
  86
  87 /* Nodes only know about those public addresses that they are
  88  * configured to serve and no individual node has a full list of all
  89  * public addresses configured across the cluster.  Therefore, a
  90  * merged list of all public addresses needs to be built so that IP
  91  * allocation can be done. */
  92 static struct public_ip_list *
  93 create_merged_ip_list(struct ipalloc_state *ipalloc_state)
  94 {
  95         unsigned int i, j;
  96         struct public_ip_list *ip_list;
  97         struct ctdb_public_ip_list *public_ips;
  98         struct trbt_tree *ip_tree;
  99         int ret;
 100
 101         ip_tree = trbt_create(ipalloc_state, 0);
 102
 103         if (ipalloc_state->known_public_ips == NULL) {
 104                 DEBUG(DEBUG_ERR, ("Known public IPs not set\n"));
 105                 return NULL;
 106         }
 107
 108         for (i=0; i < ipalloc_state->num; i++) {
 109
 110                 public_ips = &ipalloc_state->known_public_ips[i];
 111
 112                 for (j=0; j < public_ips->num; j++) {
 113                         struct public_ip_list *tmp_ip;
 114
 115                         /* This is returned as part of ip_list */
 116                         tmp_ip = talloc_zero(ipalloc_state, struct public_ip_list);
 117                         if (tmp_ip == NULL) {
 118                                 DEBUG(DEBUG_ERR,
 119                                       (__location__ " out of memory\n"));
 120                                 talloc_free(ip_tree);
 121                                 return NULL;
 122                         }
 123
 124                         /* Do not use information about IP addresses hosted
 125                          * on other nodes, it may not be accurate */
 126                         if (public_ips->ip[j].pnn == i) {
 127                                 tmp_ip->pnn = public_ips->ip[j].pnn;
 128                         } else {
 129                                 tmp_ip->pnn = CTDB_UNKNOWN_PNN;
 130                         }
 131                         tmp_ip->addr = public_ips->ip[j].addr;
 132                         tmp_ip->next = NULL;
 133
 134                         trbt_insertarray32_callback(ip_tree,
 135                                 IP_KEYLEN, ip_key(&public_ips->ip[j].addr),
 136                                 add_ip_callback,
 137                                 tmp_ip);
 138                 }
 139         }
 140
 141         ip_list = NULL;
 142         ret = trbt_traversearray32(ip_tree, IP_KEYLEN, getips_count_callback, &ip_list);
 143         if (ret != 0) {
 144                 DBG_ERR("Error traversing the IP tree.\n");
 145         }
 146
 147         talloc_free(ip_tree);
 148
 149         return ip_list;
 150 }
 151
 152 static bool populate_bitmap(struct ipalloc_state *ipalloc_state)
 153 {
 154         struct public_ip_list *ip = NULL;
 155         unsigned int i, j;
 156
 157         for (ip = ipalloc_state->all_ips; ip != NULL; ip = ip->next) {
 158
 159                 ip->known_on = bitmap_talloc(ip, ipalloc_state->num);
 160                 if (ip->known_on == NULL) {
 161                         return false;
 162                 }
 163
 164                 ip->available_on = bitmap_talloc(ip, ipalloc_state->num);
 165                 if (ip->available_on == NULL) {
 166                         return false;
 167                 }
 168
 169                 for (i = 0; i < ipalloc_state->num; i++) {
 170                         struct ctdb_public_ip_list *known =
 171                                 &ipalloc_state->known_public_ips[i];
 172                         struct ctdb_public_ip_list *avail =
 173                                 &ipalloc_state->available_public_ips[i];
 174
 175                         /* Check to see if "ip" is available on node "i" */
 176                         for (j = 0; j < avail->num; j++) {
 177                                 if (ctdb_sock_addr_same_ip(
 178                                             &ip->addr, &avail->ip[j].addr)) {
 179                                         bitmap_set(ip->available_on, i);
 180                                         break;
 181                                 }
 182                         }
 183
 184                         /* Optimisation: available => known */
 185                         if (bitmap_query(ip->available_on, i)) {
 186                                 bitmap_set(ip->known_on, i);
 187                                 continue;
 188                         }
 189
 190                         /* Check to see if "ip" is known on node "i" */
 191                         for (j = 0; j < known->num; j++) {
 192                                 if (ctdb_sock_addr_same_ip(
 193                                             &ip->addr, &known->ip[j].addr)) {
 194                                         bitmap_set(ip->known_on, i);
 195                                         break;
 196                                 }
 197                         }
 198                 }
 199         }
 200
 201         return true;
 202 }
 203
 204 void ipalloc_set_public_ips(struct ipalloc_state *ipalloc_state,
 205                             struct ctdb_public_ip_list *known_ips,
 206                             struct ctdb_public_ip_list *available_ips)
 207 {
 208         ipalloc_state->available_public_ips = available_ips;
 209         ipalloc_state->known_public_ips = known_ips;
 210 }
 211
 212 /* This can only return false if there are no available IPs *and*
 213  * there are no IP addresses currently allocated.  If the latter is
 214  * true then the cluster can clearly host IPs... just not necessarily
 215  * right now... */
 216 bool ipalloc_can_host_ips(struct ipalloc_state *ipalloc_state)
 217 {
 218         unsigned int i;
 219         bool have_ips = false;
 220
 221         for (i=0; i < ipalloc_state->num; i++) {
 222                 struct ctdb_public_ip_list *ips =
 223                         ipalloc_state->known_public_ips;
 224                 if (ips[i].num != 0) {
 225                         unsigned int j;
 226                         have_ips = true;
 227                         /* Succeed if an address is hosted on node i */
 228                         for (j=0; j < ips[i].num; j++) {
 229                                 if (ips[i].ip[j].pnn == i) {
 230                                         return true;
 231                                 }
 232                         }
 233                 }
 234         }
 235
 236         if (! have_ips) {
 237                 return false;
 238         }
 239
 240         /* At this point there are known addresses but none are
 241          * hosted.  Need to check if cluster can now host some
 242          * addresses.
 243          */
 244         for (i=0; i < ipalloc_state->num; i++) {
 245                 if (ipalloc_state->available_public_ips[i].num != 0) {
 246                         return true;
 247                 }
 248         }
 249
 250         return false;
 251 }
 252
 253 /* The calculation part of the IP allocation algorithm. */
 254 struct public_ip_list *ipalloc(struct ipalloc_state *ipalloc_state)
 255 {
 256         bool ret = false;
 257
 258         ipalloc_state->all_ips = create_merged_ip_list(ipalloc_state);
 259         if (ipalloc_state->all_ips == NULL) {
 260                 return NULL;
 261         }
 262
 263         if (!populate_bitmap(ipalloc_state)) {
 264                 return NULL;
 265         }
 266
 267         switch (ipalloc_state->algorithm) {
 268         case IPALLOC_LCP2:
 269                 ret = ipalloc_lcp2(ipalloc_state);
 270                 break;
 271         case IPALLOC_DETERMINISTIC:
 272                 ret = ipalloc_deterministic(ipalloc_state);
 273                 break;
 274         case IPALLOC_NONDETERMINISTIC:
 275                 ret = ipalloc_nondeterministic(ipalloc_state);
 276                break;
 277         }
 278
 279         /* at this point ->pnn is the node which will own each IP
 280            or CTDB_UNKNOWN_PNN if there is no node that can cover this ip
 281         */
 282
 283         return (ret ? ipalloc_state->all_ips : NULL);
 284 }