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/>.
22 #include "system/network.h"
23 #include "system/filesys.h"
24 #include "system/time.h"
25 #include "system/wait.h"
30 #include "lib/util/dlinklist.h"
31 #include "lib/util/debug.h"
32 #include "lib/util/samba_util.h"
33 #include "lib/util/util_process.h"
35 #include "ctdb_private.h"
36 #include "ctdb_client.h"
38 #include "common/rb_tree.h"
39 #include "common/reqid.h"
40 #include "common/system.h"
41 #include "common/common.h"
42 #include "common/logging.h"
45 #define TAKEOVER_TIMEOUT() timeval_current_ofs(ctdb->tunable.takeover_timeout,0)
47 #define CTDB_ARP_INTERVAL 1
48 #define CTDB_ARP_REPEAT 3
50 /* Flags used in IP allocation algorithms. */
56 enum ipalloc_algorithm {
57 IPALLOC_DETERMINISTIC,
58 IPALLOC_NONDETERMINISTIC,
62 struct ipalloc_state {
65 /* Arrays with data for each node */
66 struct ctdb_public_ip_list_old **known_public_ips;
67 struct ctdb_public_ip_list_old **available_public_ips;
68 struct ctdb_ipflags *ipflags;
70 enum ipalloc_algorithm algorithm;
71 uint32_t no_ip_failback;
74 struct ctdb_interface {
75 struct ctdb_interface *prev, *next;
81 static const char *ctdb_vnn_iface_string(const struct ctdb_vnn *vnn)
84 return vnn->iface->name;
90 static int ctdb_add_local_iface(struct ctdb_context *ctdb, const char *iface)
92 struct ctdb_interface *i;
94 /* Verify that we don't have an entry for this ip yet */
95 for (i=ctdb->ifaces;i;i=i->next) {
96 if (strcmp(i->name, iface) == 0) {
101 /* create a new structure for this interface */
102 i = talloc_zero(ctdb, struct ctdb_interface);
103 CTDB_NO_MEMORY_FATAL(ctdb, i);
104 i->name = talloc_strdup(i, iface);
105 CTDB_NO_MEMORY(ctdb, i->name);
109 DLIST_ADD(ctdb->ifaces, i);
114 static bool vnn_has_interface_with_name(struct ctdb_vnn *vnn,
119 for (n = 0; vnn->ifaces[n] != NULL; n++) {
120 if (strcmp(name, vnn->ifaces[n]) == 0) {
128 /* If any interfaces now have no possible IPs then delete them. This
129 * implementation is naive (i.e. simple) rather than clever
130 * (i.e. complex). Given that this is run on delip and that operation
131 * is rare, this doesn't need to be efficient - it needs to be
132 * foolproof. One alternative is reference counting, where the logic
133 * is distributed and can, therefore, be broken in multiple places.
134 * Another alternative is to build a red-black tree of interfaces that
135 * can have addresses (by walking ctdb->vnn and ctdb->single_ip_vnn
136 * once) and then walking ctdb->ifaces once and deleting those not in
137 * the tree. Let's go to one of those if the naive implementation
138 * causes problems... :-)
140 static void ctdb_remove_orphaned_ifaces(struct ctdb_context *ctdb,
141 struct ctdb_vnn *vnn)
143 struct ctdb_interface *i, *next;
145 /* For each interface, check if there's an IP using it. */
146 for (i = ctdb->ifaces; i != NULL; i = next) {
151 /* Only consider interfaces named in the given VNN. */
152 if (!vnn_has_interface_with_name(vnn, i->name)) {
156 /* Is the "single IP" on this interface? */
157 if ((ctdb->single_ip_vnn != NULL) &&
158 (ctdb->single_ip_vnn->ifaces[0] != NULL) &&
159 (strcmp(i->name, ctdb->single_ip_vnn->ifaces[0]) == 0)) {
160 /* Found, next interface please... */
163 /* Search for a vnn with this interface. */
165 for (tv=ctdb->vnn; tv; tv=tv->next) {
166 if (vnn_has_interface_with_name(tv, i->name)) {
173 /* None of the VNNs are using this interface. */
174 DLIST_REMOVE(ctdb->ifaces, i);
181 static struct ctdb_interface *ctdb_find_iface(struct ctdb_context *ctdb,
184 struct ctdb_interface *i;
186 for (i=ctdb->ifaces;i;i=i->next) {
187 if (strcmp(i->name, iface) == 0) {
195 static struct ctdb_interface *ctdb_vnn_best_iface(struct ctdb_context *ctdb,
196 struct ctdb_vnn *vnn)
199 struct ctdb_interface *cur = NULL;
200 struct ctdb_interface *best = NULL;
202 for (i=0; vnn->ifaces[i]; i++) {
204 cur = ctdb_find_iface(ctdb, vnn->ifaces[i]);
218 if (cur->references < best->references) {
227 static int32_t ctdb_vnn_assign_iface(struct ctdb_context *ctdb,
228 struct ctdb_vnn *vnn)
230 struct ctdb_interface *best = NULL;
233 DEBUG(DEBUG_INFO, (__location__ " public address '%s' "
234 "still assigned to iface '%s'\n",
235 ctdb_addr_to_str(&vnn->public_address),
236 ctdb_vnn_iface_string(vnn)));
240 best = ctdb_vnn_best_iface(ctdb, vnn);
242 DEBUG(DEBUG_ERR, (__location__ " public address '%s' "
243 "cannot assign to iface any iface\n",
244 ctdb_addr_to_str(&vnn->public_address)));
250 vnn->pnn = ctdb->pnn;
252 DEBUG(DEBUG_INFO, (__location__ " public address '%s' "
253 "now assigned to iface '%s' refs[%d]\n",
254 ctdb_addr_to_str(&vnn->public_address),
255 ctdb_vnn_iface_string(vnn),
260 static void ctdb_vnn_unassign_iface(struct ctdb_context *ctdb,
261 struct ctdb_vnn *vnn)
263 DEBUG(DEBUG_INFO, (__location__ " public address '%s' "
264 "now unassigned (old iface '%s' refs[%d])\n",
265 ctdb_addr_to_str(&vnn->public_address),
266 ctdb_vnn_iface_string(vnn),
267 vnn->iface?vnn->iface->references:0));
269 vnn->iface->references--;
272 if (vnn->pnn == ctdb->pnn) {
277 static bool ctdb_vnn_available(struct ctdb_context *ctdb,
278 struct ctdb_vnn *vnn)
282 /* Nodes that are not RUNNING can not host IPs */
283 if (ctdb->runstate != CTDB_RUNSTATE_RUNNING) {
287 if (vnn->delete_pending) {
291 if (vnn->iface && vnn->iface->link_up) {
295 for (i=0; vnn->ifaces[i]; i++) {
296 struct ctdb_interface *cur;
298 cur = ctdb_find_iface(ctdb, vnn->ifaces[i]);
311 struct ctdb_takeover_arp {
312 struct ctdb_context *ctdb;
315 struct ctdb_tcp_array *tcparray;
316 struct ctdb_vnn *vnn;
321 lists of tcp endpoints
323 struct ctdb_tcp_list {
324 struct ctdb_tcp_list *prev, *next;
325 struct ctdb_connection connection;
329 list of clients to kill on IP release
331 struct ctdb_client_ip {
332 struct ctdb_client_ip *prev, *next;
333 struct ctdb_context *ctdb;
340 send a gratuitous arp
342 static void ctdb_control_send_arp(struct tevent_context *ev,
343 struct tevent_timer *te,
344 struct timeval t, void *private_data)
346 struct ctdb_takeover_arp *arp = talloc_get_type(private_data,
347 struct ctdb_takeover_arp);
349 struct ctdb_tcp_array *tcparray;
350 const char *iface = ctdb_vnn_iface_string(arp->vnn);
352 ret = ctdb_sys_send_arp(&arp->addr, iface);
354 DEBUG(DEBUG_CRIT,(__location__ " sending of arp failed on iface '%s' (%s)\n",
355 iface, strerror(errno)));
358 tcparray = arp->tcparray;
360 for (i=0;i<tcparray->num;i++) {
361 struct ctdb_connection *tcon;
363 tcon = &tcparray->connections[i];
364 DEBUG(DEBUG_INFO,("sending tcp tickle ack for %u->%s:%u\n",
365 (unsigned)ntohs(tcon->dst.ip.sin_port),
366 ctdb_addr_to_str(&tcon->src),
367 (unsigned)ntohs(tcon->src.ip.sin_port)));
368 ret = ctdb_sys_send_tcp(
373 DEBUG(DEBUG_CRIT,(__location__ " Failed to send tcp tickle ack for %s\n",
374 ctdb_addr_to_str(&tcon->src)));
381 if (arp->count == CTDB_ARP_REPEAT) {
386 tevent_add_timer(arp->ctdb->ev, arp->vnn->takeover_ctx,
387 timeval_current_ofs(CTDB_ARP_INTERVAL, 100000),
388 ctdb_control_send_arp, arp);
391 static int32_t ctdb_announce_vnn_iface(struct ctdb_context *ctdb,
392 struct ctdb_vnn *vnn)
394 struct ctdb_takeover_arp *arp;
395 struct ctdb_tcp_array *tcparray;
397 if (!vnn->takeover_ctx) {
398 vnn->takeover_ctx = talloc_new(vnn);
399 if (!vnn->takeover_ctx) {
404 arp = talloc_zero(vnn->takeover_ctx, struct ctdb_takeover_arp);
410 arp->addr = vnn->public_address;
413 tcparray = vnn->tcp_array;
415 /* add all of the known tcp connections for this IP to the
416 list of tcp connections to send tickle acks for */
417 arp->tcparray = talloc_steal(arp, tcparray);
419 vnn->tcp_array = NULL;
420 vnn->tcp_update_needed = true;
423 tevent_add_timer(arp->ctdb->ev, vnn->takeover_ctx,
424 timeval_zero(), ctdb_control_send_arp, arp);
429 struct takeover_callback_state {
430 struct ctdb_req_control_old *c;
431 ctdb_sock_addr *addr;
432 struct ctdb_vnn *vnn;
435 struct ctdb_do_takeip_state {
436 struct ctdb_req_control_old *c;
437 struct ctdb_vnn *vnn;
441 called when takeip event finishes
443 static void ctdb_do_takeip_callback(struct ctdb_context *ctdb, int status,
446 struct ctdb_do_takeip_state *state =
447 talloc_get_type(private_data, struct ctdb_do_takeip_state);
452 struct ctdb_node *node = ctdb->nodes[ctdb->pnn];
454 if (status == -ETIME) {
457 DEBUG(DEBUG_ERR,(__location__ " Failed to takeover IP %s on interface %s\n",
458 ctdb_addr_to_str(&state->vnn->public_address),
459 ctdb_vnn_iface_string(state->vnn)));
460 ctdb_request_control_reply(ctdb, state->c, NULL, status, NULL);
462 node->flags |= NODE_FLAGS_UNHEALTHY;
467 if (ctdb->do_checkpublicip) {
469 ret = ctdb_announce_vnn_iface(ctdb, state->vnn);
471 ctdb_request_control_reply(ctdb, state->c, NULL, -1, NULL);
478 data.dptr = (uint8_t *)ctdb_addr_to_str(&state->vnn->public_address);
479 data.dsize = strlen((char *)data.dptr) + 1;
480 DEBUG(DEBUG_INFO,(__location__ " sending TAKE_IP for '%s'\n", data.dptr));
482 ctdb_daemon_send_message(ctdb, ctdb->pnn, CTDB_SRVID_TAKE_IP, data);
485 /* the control succeeded */
486 ctdb_request_control_reply(ctdb, state->c, NULL, 0, NULL);
491 static int ctdb_takeip_destructor(struct ctdb_do_takeip_state *state)
493 state->vnn->update_in_flight = false;
498 take over an ip address
500 static int32_t ctdb_do_takeip(struct ctdb_context *ctdb,
501 struct ctdb_req_control_old *c,
502 struct ctdb_vnn *vnn)
505 struct ctdb_do_takeip_state *state;
507 if (vnn->update_in_flight) {
508 DEBUG(DEBUG_NOTICE,("Takeover of IP %s/%u rejected "
509 "update for this IP already in flight\n",
510 ctdb_addr_to_str(&vnn->public_address),
511 vnn->public_netmask_bits));
515 ret = ctdb_vnn_assign_iface(ctdb, vnn);
517 DEBUG(DEBUG_ERR,("Takeover of IP %s/%u failed to "
518 "assign a usable interface\n",
519 ctdb_addr_to_str(&vnn->public_address),
520 vnn->public_netmask_bits));
524 state = talloc(vnn, struct ctdb_do_takeip_state);
525 CTDB_NO_MEMORY(ctdb, state);
527 state->c = talloc_steal(ctdb, c);
530 vnn->update_in_flight = true;
531 talloc_set_destructor(state, ctdb_takeip_destructor);
533 DEBUG(DEBUG_NOTICE,("Takeover of IP %s/%u on interface %s\n",
534 ctdb_addr_to_str(&vnn->public_address),
535 vnn->public_netmask_bits,
536 ctdb_vnn_iface_string(vnn)));
538 ret = ctdb_event_script_callback(ctdb,
540 ctdb_do_takeip_callback,
544 ctdb_vnn_iface_string(vnn),
545 ctdb_addr_to_str(&vnn->public_address),
546 vnn->public_netmask_bits);
549 DEBUG(DEBUG_ERR,(__location__ " Failed to takeover IP %s on interface %s\n",
550 ctdb_addr_to_str(&vnn->public_address),
551 ctdb_vnn_iface_string(vnn)));
559 struct ctdb_do_updateip_state {
560 struct ctdb_req_control_old *c;
561 struct ctdb_interface *old;
562 struct ctdb_vnn *vnn;
566 called when updateip event finishes
568 static void ctdb_do_updateip_callback(struct ctdb_context *ctdb, int status,
571 struct ctdb_do_updateip_state *state =
572 talloc_get_type(private_data, struct ctdb_do_updateip_state);
576 if (status == -ETIME) {
579 DEBUG(DEBUG_ERR,(__location__ " Failed to move IP %s from interface %s to %s\n",
580 ctdb_addr_to_str(&state->vnn->public_address),
582 ctdb_vnn_iface_string(state->vnn)));
585 * All we can do is reset the old interface
586 * and let the next run fix it
588 ctdb_vnn_unassign_iface(ctdb, state->vnn);
589 state->vnn->iface = state->old;
590 state->vnn->iface->references++;
592 ctdb_request_control_reply(ctdb, state->c, NULL, status, NULL);
597 if (ctdb->do_checkpublicip) {
599 ret = ctdb_announce_vnn_iface(ctdb, state->vnn);
601 ctdb_request_control_reply(ctdb, state->c, NULL, -1, NULL);
608 /* the control succeeded */
609 ctdb_request_control_reply(ctdb, state->c, NULL, 0, NULL);
614 static int ctdb_updateip_destructor(struct ctdb_do_updateip_state *state)
616 state->vnn->update_in_flight = false;
621 update (move) an ip address
623 static int32_t ctdb_do_updateip(struct ctdb_context *ctdb,
624 struct ctdb_req_control_old *c,
625 struct ctdb_vnn *vnn)
628 struct ctdb_do_updateip_state *state;
629 struct ctdb_interface *old = vnn->iface;
630 const char *new_name;
632 if (vnn->update_in_flight) {
633 DEBUG(DEBUG_NOTICE,("Update of IP %s/%u rejected "
634 "update for this IP already in flight\n",
635 ctdb_addr_to_str(&vnn->public_address),
636 vnn->public_netmask_bits));
640 ctdb_vnn_unassign_iface(ctdb, vnn);
641 ret = ctdb_vnn_assign_iface(ctdb, vnn);
643 DEBUG(DEBUG_ERR,("update of IP %s/%u failed to "
644 "assin a usable interface (old iface '%s')\n",
645 ctdb_addr_to_str(&vnn->public_address),
646 vnn->public_netmask_bits,
651 new_name = ctdb_vnn_iface_string(vnn);
652 if (old->name != NULL && new_name != NULL && !strcmp(old->name, new_name)) {
653 /* A benign update from one interface onto itself.
654 * no need to run the eventscripts in this case, just return
657 ctdb_request_control_reply(ctdb, c, NULL, 0, NULL);
661 state = talloc(vnn, struct ctdb_do_updateip_state);
662 CTDB_NO_MEMORY(ctdb, state);
664 state->c = talloc_steal(ctdb, c);
668 vnn->update_in_flight = true;
669 talloc_set_destructor(state, ctdb_updateip_destructor);
671 DEBUG(DEBUG_NOTICE,("Update of IP %s/%u from "
672 "interface %s to %s\n",
673 ctdb_addr_to_str(&vnn->public_address),
674 vnn->public_netmask_bits,
678 ret = ctdb_event_script_callback(ctdb,
680 ctdb_do_updateip_callback,
682 CTDB_EVENT_UPDATE_IP,
686 ctdb_addr_to_str(&vnn->public_address),
687 vnn->public_netmask_bits);
689 DEBUG(DEBUG_ERR,(__location__ " Failed update IP %s from interface %s to %s\n",
690 ctdb_addr_to_str(&vnn->public_address),
691 old->name, new_name));
700 Find the vnn of the node that has a public ip address
701 returns -1 if the address is not known as a public address
703 static struct ctdb_vnn *find_public_ip_vnn(struct ctdb_context *ctdb, ctdb_sock_addr *addr)
705 struct ctdb_vnn *vnn;
707 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
708 if (ctdb_same_ip(&vnn->public_address, addr)) {
717 take over an ip address
719 int32_t ctdb_control_takeover_ip(struct ctdb_context *ctdb,
720 struct ctdb_req_control_old *c,
725 struct ctdb_public_ip *pip = (struct ctdb_public_ip *)indata.dptr;
726 struct ctdb_vnn *vnn;
727 bool have_ip = false;
728 bool do_updateip = false;
729 bool do_takeip = false;
730 struct ctdb_interface *best_iface = NULL;
732 if (pip->pnn != ctdb->pnn) {
733 DEBUG(DEBUG_ERR,(__location__" takeoverip called for an ip '%s' "
734 "with pnn %d, but we're node %d\n",
735 ctdb_addr_to_str(&pip->addr),
736 pip->pnn, ctdb->pnn));
740 /* update out vnn list */
741 vnn = find_public_ip_vnn(ctdb, &pip->addr);
743 DEBUG(DEBUG_INFO,("takeoverip called for an ip '%s' that is not a public address\n",
744 ctdb_addr_to_str(&pip->addr)));
748 if (ctdb->tunable.disable_ip_failover == 0 && ctdb->do_checkpublicip) {
749 have_ip = ctdb_sys_have_ip(&pip->addr);
751 best_iface = ctdb_vnn_best_iface(ctdb, vnn);
752 if (best_iface == NULL) {
753 DEBUG(DEBUG_ERR,("takeoverip of IP %s/%u failed to find"
754 "a usable interface (old %s, have_ip %d)\n",
755 ctdb_addr_to_str(&vnn->public_address),
756 vnn->public_netmask_bits,
757 ctdb_vnn_iface_string(vnn),
762 if (vnn->iface == NULL && vnn->pnn == -1 && have_ip && best_iface != NULL) {
763 DEBUG(DEBUG_ERR,("Taking over newly created ip\n"));
768 if (vnn->iface == NULL && have_ip) {
769 DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
770 "but we have no interface assigned, has someone manually configured it? Ignore for now.\n",
771 ctdb_addr_to_str(&vnn->public_address)));
775 if (vnn->pnn != ctdb->pnn && have_ip && vnn->pnn != -1) {
776 DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
777 "and we have it on iface[%s], but it was assigned to node %d"
778 "and we are node %d, banning ourself\n",
779 ctdb_addr_to_str(&vnn->public_address),
780 ctdb_vnn_iface_string(vnn), vnn->pnn, ctdb->pnn));
785 if (vnn->pnn == -1 && have_ip) {
786 vnn->pnn = ctdb->pnn;
787 DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
788 "and we already have it on iface[%s], update local daemon\n",
789 ctdb_addr_to_str(&vnn->public_address),
790 ctdb_vnn_iface_string(vnn)));
795 if (vnn->iface != best_iface) {
796 if (!vnn->iface->link_up) {
798 } else if (vnn->iface->references > (best_iface->references + 1)) {
799 /* only move when the rebalance gains something */
807 ctdb_vnn_unassign_iface(ctdb, vnn);
814 ret = ctdb_do_takeip(ctdb, c, vnn);
818 } else if (do_updateip) {
819 ret = ctdb_do_updateip(ctdb, c, vnn);
825 * The interface is up and the kernel known the ip
828 DEBUG(DEBUG_INFO,("Redundant takeover of IP %s/%u on interface %s (ip already held)\n",
829 ctdb_addr_to_str(&pip->addr),
830 vnn->public_netmask_bits,
831 ctdb_vnn_iface_string(vnn)));
835 /* tell ctdb_control.c that we will be replying asynchronously */
842 kill any clients that are registered with a IP that is being released
844 static void release_kill_clients(struct ctdb_context *ctdb, ctdb_sock_addr *addr)
846 struct ctdb_client_ip *ip;
848 DEBUG(DEBUG_INFO,("release_kill_clients for ip %s\n",
849 ctdb_addr_to_str(addr)));
851 for (ip=ctdb->client_ip_list; ip; ip=ip->next) {
852 ctdb_sock_addr tmp_addr;
855 DEBUG(DEBUG_INFO,("checking for client %u with IP %s\n",
857 ctdb_addr_to_str(&ip->addr)));
859 if (ctdb_same_ip(&tmp_addr, addr)) {
860 struct ctdb_client *client = reqid_find(ctdb->idr,
863 DEBUG(DEBUG_INFO,("matched client %u with IP %s and pid %u\n",
865 ctdb_addr_to_str(&ip->addr),
868 if (client->pid != 0) {
869 DEBUG(DEBUG_INFO,(__location__ " Killing client pid %u for IP %s on client_id %u\n",
870 (unsigned)client->pid,
871 ctdb_addr_to_str(addr),
873 kill(client->pid, SIGKILL);
879 static void do_delete_ip(struct ctdb_context *ctdb, struct ctdb_vnn *vnn)
881 DLIST_REMOVE(ctdb->vnn, vnn);
882 ctdb_vnn_unassign_iface(ctdb, vnn);
883 ctdb_remove_orphaned_ifaces(ctdb, vnn);
888 called when releaseip event finishes
890 static void release_ip_callback(struct ctdb_context *ctdb, int status,
893 struct takeover_callback_state *state =
894 talloc_get_type(private_data, struct takeover_callback_state);
897 if (status == -ETIME) {
901 if (ctdb->tunable.disable_ip_failover == 0 && ctdb->do_checkpublicip) {
902 if (ctdb_sys_have_ip(state->addr)) {
904 ("IP %s still hosted during release IP callback, failing\n",
905 ctdb_addr_to_str(state->addr)));
906 ctdb_request_control_reply(ctdb, state->c,
913 /* send a message to all clients of this node telling them
914 that the cluster has been reconfigured and they should
915 release any sockets on this IP */
916 data.dptr = (uint8_t *)talloc_strdup(state, ctdb_addr_to_str(state->addr));
917 CTDB_NO_MEMORY_VOID(ctdb, data.dptr);
918 data.dsize = strlen((char *)data.dptr)+1;
920 DEBUG(DEBUG_INFO,(__location__ " sending RELEASE_IP for '%s'\n", data.dptr));
922 ctdb_daemon_send_message(ctdb, ctdb->pnn, CTDB_SRVID_RELEASE_IP, data);
924 /* kill clients that have registered with this IP */
925 release_kill_clients(ctdb, state->addr);
927 ctdb_vnn_unassign_iface(ctdb, state->vnn);
929 /* Process the IP if it has been marked for deletion */
930 if (state->vnn->delete_pending) {
931 do_delete_ip(ctdb, state->vnn);
935 /* the control succeeded */
936 ctdb_request_control_reply(ctdb, state->c, NULL, 0, NULL);
940 static int ctdb_releaseip_destructor(struct takeover_callback_state *state)
942 if (state->vnn != NULL) {
943 state->vnn->update_in_flight = false;
949 release an ip address
951 int32_t ctdb_control_release_ip(struct ctdb_context *ctdb,
952 struct ctdb_req_control_old *c,
957 struct takeover_callback_state *state;
958 struct ctdb_public_ip *pip = (struct ctdb_public_ip *)indata.dptr;
959 struct ctdb_vnn *vnn;
962 /* update our vnn list */
963 vnn = find_public_ip_vnn(ctdb, &pip->addr);
965 DEBUG(DEBUG_INFO,("releaseip called for an ip '%s' that is not a public address\n",
966 ctdb_addr_to_str(&pip->addr)));
971 /* stop any previous arps */
972 talloc_free(vnn->takeover_ctx);
973 vnn->takeover_ctx = NULL;
975 /* Some ctdb tool commands (e.g. moveip, rebalanceip) send
976 * lazy multicast to drop an IP from any node that isn't the
977 * intended new node. The following causes makes ctdbd ignore
978 * a release for any address it doesn't host.
980 if (ctdb->tunable.disable_ip_failover == 0 && ctdb->do_checkpublicip) {
981 if (!ctdb_sys_have_ip(&pip->addr)) {
982 DEBUG(DEBUG_DEBUG,("Redundant release of IP %s/%u on interface %s (ip not held)\n",
983 ctdb_addr_to_str(&pip->addr),
984 vnn->public_netmask_bits,
985 ctdb_vnn_iface_string(vnn)));
986 ctdb_vnn_unassign_iface(ctdb, vnn);
990 if (vnn->iface == NULL) {
991 DEBUG(DEBUG_DEBUG,("Redundant release of IP %s/%u (ip not held)\n",
992 ctdb_addr_to_str(&pip->addr),
993 vnn->public_netmask_bits));
998 /* There is a potential race between take_ip and us because we
999 * update the VNN via a callback that run when the
1000 * eventscripts have been run. Avoid the race by allowing one
1001 * update to be in flight at a time.
1003 if (vnn->update_in_flight) {
1004 DEBUG(DEBUG_NOTICE,("Release of IP %s/%u rejected "
1005 "update for this IP already in flight\n",
1006 ctdb_addr_to_str(&vnn->public_address),
1007 vnn->public_netmask_bits));
1011 iface = strdup(ctdb_vnn_iface_string(vnn));
1013 DEBUG(DEBUG_NOTICE,("Release of IP %s/%u on interface %s node:%d\n",
1014 ctdb_addr_to_str(&pip->addr),
1015 vnn->public_netmask_bits,
1019 state = talloc(ctdb, struct takeover_callback_state);
1020 if (state == NULL) {
1021 ctdb_set_error(ctdb, "Out of memory at %s:%d",
1022 __FILE__, __LINE__);
1027 state->c = talloc_steal(state, c);
1028 state->addr = talloc(state, ctdb_sock_addr);
1029 if (state->addr == NULL) {
1030 ctdb_set_error(ctdb, "Out of memory at %s:%d",
1031 __FILE__, __LINE__);
1036 *state->addr = pip->addr;
1039 vnn->update_in_flight = true;
1040 talloc_set_destructor(state, ctdb_releaseip_destructor);
1042 ret = ctdb_event_script_callback(ctdb,
1043 state, release_ip_callback, state,
1044 CTDB_EVENT_RELEASE_IP,
1047 ctdb_addr_to_str(&pip->addr),
1048 vnn->public_netmask_bits);
1051 DEBUG(DEBUG_ERR,(__location__ " Failed to release IP %s on interface %s\n",
1052 ctdb_addr_to_str(&pip->addr),
1053 ctdb_vnn_iface_string(vnn)));
1058 /* tell the control that we will be reply asynchronously */
1059 *async_reply = true;
1063 static int ctdb_add_public_address(struct ctdb_context *ctdb,
1064 ctdb_sock_addr *addr,
1065 unsigned mask, const char *ifaces,
1068 struct ctdb_vnn *vnn;
1075 tmp = strdup(ifaces);
1076 for (iface = strtok(tmp, ","); iface; iface = strtok(NULL, ",")) {
1077 if (!ctdb_sys_check_iface_exists(iface)) {
1078 DEBUG(DEBUG_CRIT,("Interface %s does not exist. Can not add public-address : %s\n", iface, ctdb_addr_to_str(addr)));
1085 /* Verify that we don't have an entry for this ip yet */
1086 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
1087 if (ctdb_same_sockaddr(addr, &vnn->public_address)) {
1088 DEBUG(DEBUG_CRIT,("Same ip '%s' specified multiple times in the public address list \n",
1089 ctdb_addr_to_str(addr)));
1094 /* create a new vnn structure for this ip address */
1095 vnn = talloc_zero(ctdb, struct ctdb_vnn);
1096 CTDB_NO_MEMORY_FATAL(ctdb, vnn);
1097 vnn->ifaces = talloc_array(vnn, const char *, num + 2);
1098 tmp = talloc_strdup(vnn, ifaces);
1099 CTDB_NO_MEMORY_FATAL(ctdb, tmp);
1100 for (iface = strtok(tmp, ","); iface; iface = strtok(NULL, ",")) {
1101 vnn->ifaces = talloc_realloc(vnn, vnn->ifaces, const char *, num + 2);
1102 CTDB_NO_MEMORY_FATAL(ctdb, vnn->ifaces);
1103 vnn->ifaces[num] = talloc_strdup(vnn, iface);
1104 CTDB_NO_MEMORY_FATAL(ctdb, vnn->ifaces[num]);
1108 vnn->ifaces[num] = NULL;
1109 vnn->public_address = *addr;
1110 vnn->public_netmask_bits = mask;
1112 if (check_address) {
1113 if (ctdb_sys_have_ip(addr)) {
1114 DEBUG(DEBUG_ERR,("We are already hosting public address '%s'. setting PNN to ourself:%d\n", ctdb_addr_to_str(addr), ctdb->pnn));
1115 vnn->pnn = ctdb->pnn;
1119 for (i=0; vnn->ifaces[i]; i++) {
1120 ret = ctdb_add_local_iface(ctdb, vnn->ifaces[i]);
1122 DEBUG(DEBUG_CRIT, (__location__ " failed to add iface[%s] "
1123 "for public_address[%s]\n",
1124 vnn->ifaces[i], ctdb_addr_to_str(addr)));
1130 DLIST_ADD(ctdb->vnn, vnn);
1136 setup the public address lists from a file
1138 int ctdb_set_public_addresses(struct ctdb_context *ctdb, bool check_addresses)
1144 lines = file_lines_load(ctdb->public_addresses_file, &nlines, 0, ctdb);
1145 if (lines == NULL) {
1146 ctdb_set_error(ctdb, "Failed to load public address list '%s'\n", ctdb->public_addresses_file);
1149 while (nlines > 0 && strcmp(lines[nlines-1], "") == 0) {
1153 for (i=0;i<nlines;i++) {
1155 ctdb_sock_addr addr;
1156 const char *addrstr;
1161 while ((*line == ' ') || (*line == '\t')) {
1167 if (strcmp(line, "") == 0) {
1170 tok = strtok(line, " \t");
1172 tok = strtok(NULL, " \t");
1174 if (NULL == ctdb->default_public_interface) {
1175 DEBUG(DEBUG_CRIT,("No default public interface and no interface specified at line %u of public address list\n",
1180 ifaces = ctdb->default_public_interface;
1185 if (!addrstr || !parse_ip_mask(addrstr, ifaces, &addr, &mask)) {
1186 DEBUG(DEBUG_CRIT,("Badly formed line %u in public address list\n", i+1));
1190 if (ctdb_add_public_address(ctdb, &addr, mask, ifaces, check_addresses)) {
1191 DEBUG(DEBUG_CRIT,("Failed to add line %u to the public address list\n", i+1));
1202 int ctdb_set_single_public_ip(struct ctdb_context *ctdb,
1206 struct ctdb_vnn *svnn;
1207 struct ctdb_interface *cur = NULL;
1211 svnn = talloc_zero(ctdb, struct ctdb_vnn);
1212 CTDB_NO_MEMORY(ctdb, svnn);
1214 svnn->ifaces = talloc_array(svnn, const char *, 2);
1215 CTDB_NO_MEMORY(ctdb, svnn->ifaces);
1216 svnn->ifaces[0] = talloc_strdup(svnn->ifaces, iface);
1217 CTDB_NO_MEMORY(ctdb, svnn->ifaces[0]);
1218 svnn->ifaces[1] = NULL;
1220 ok = parse_ip(ip, iface, 0, &svnn->public_address);
1226 ret = ctdb_add_local_iface(ctdb, svnn->ifaces[0]);
1228 DEBUG(DEBUG_CRIT, (__location__ " failed to add iface[%s] "
1229 "for single_ip[%s]\n",
1231 ctdb_addr_to_str(&svnn->public_address)));
1236 /* assume the single public ip interface is initially "good" */
1237 cur = ctdb_find_iface(ctdb, iface);
1239 DEBUG(DEBUG_CRIT,("Can not find public interface %s used by --single-public-ip", iface));
1242 cur->link_up = true;
1244 ret = ctdb_vnn_assign_iface(ctdb, svnn);
1250 ctdb->single_ip_vnn = svnn;
1254 struct public_ip_list {
1255 struct public_ip_list *next;
1257 ctdb_sock_addr addr;
1260 /* Given a physical node, return the number of
1261 public addresses that is currently assigned to this node.
1263 static int node_ip_coverage(int32_t pnn, struct public_ip_list *ips)
1267 for (;ips;ips=ips->next) {
1268 if (ips->pnn == pnn) {
1276 /* Can the given node host the given IP: is the public IP known to the
1277 * node and is NOIPHOST unset?
1279 static bool can_node_host_ip(struct ipalloc_state *ipalloc_state,
1281 struct public_ip_list *ip)
1283 struct ctdb_public_ip_list_old *public_ips;
1286 if (ipalloc_state->ipflags[pnn].noiphost) {
1290 public_ips = ipalloc_state->available_public_ips[pnn];
1292 if (public_ips == NULL) {
1296 for (i=0; i<public_ips->num; i++) {
1297 if (ctdb_same_ip(&ip->addr, &public_ips->ips[i].addr)) {
1298 /* yes, this node can serve this public ip */
1306 static bool can_node_takeover_ip(struct ipalloc_state *ipalloc_state,
1308 struct public_ip_list *ip)
1310 if (ipalloc_state->ipflags[pnn].noiptakeover) {
1314 return can_node_host_ip(ipalloc_state, pnn, ip);
1317 /* search the node lists list for a node to takeover this ip.
1318 pick the node that currently are serving the least number of ips
1319 so that the ips get spread out evenly.
1321 static int find_takeover_node(struct ipalloc_state *ipalloc_state,
1322 struct public_ip_list *ip,
1323 struct public_ip_list *all_ips)
1325 int pnn, min=0, num;
1328 numnodes = ipalloc_state->num;
1330 for (i=0; i<numnodes; i++) {
1331 /* verify that this node can serve this ip */
1332 if (!can_node_takeover_ip(ipalloc_state, i, ip)) {
1333 /* no it couldnt so skip to the next node */
1337 num = node_ip_coverage(i, all_ips);
1338 /* was this the first node we checked ? */
1350 DEBUG(DEBUG_WARNING,(__location__ " Could not find node to take over public address '%s'\n",
1351 ctdb_addr_to_str(&ip->addr)));
1361 static uint32_t *ip_key(ctdb_sock_addr *ip)
1363 static uint32_t key[IP_KEYLEN];
1365 bzero(key, sizeof(key));
1367 switch (ip->sa.sa_family) {
1369 key[3] = htonl(ip->ip.sin_addr.s_addr);
1372 uint32_t *s6_a32 = (uint32_t *)&(ip->ip6.sin6_addr.s6_addr);
1373 key[0] = htonl(s6_a32[0]);
1374 key[1] = htonl(s6_a32[1]);
1375 key[2] = htonl(s6_a32[2]);
1376 key[3] = htonl(s6_a32[3]);
1380 DEBUG(DEBUG_ERR, (__location__ " ERROR, unknown family passed :%u\n", ip->sa.sa_family));
1387 static void *add_ip_callback(void *parm, void *data)
1389 struct public_ip_list *this_ip = parm;
1390 struct public_ip_list *prev_ip = data;
1392 if (prev_ip == NULL) {
1395 if (this_ip->pnn == -1) {
1396 this_ip->pnn = prev_ip->pnn;
1402 static int getips_count_callback(void *param, void *data)
1404 struct public_ip_list **ip_list = (struct public_ip_list **)param;
1405 struct public_ip_list *new_ip = (struct public_ip_list *)data;
1407 new_ip->next = *ip_list;
1412 static int verify_remote_ip_allocation(struct ctdb_context *ctdb,
1413 struct ctdb_public_ip_list_old *ips,
1416 static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
1417 struct ipalloc_state *ipalloc_state,
1418 struct ctdb_node_map_old *nodemap)
1423 if (ipalloc_state->num != nodemap->num) {
1426 " ipalloc_state->num (%d) != nodemap->num (%d) invalid param\n",
1427 ipalloc_state->num, nodemap->num));
1431 for (j=0; j<nodemap->num; j++) {
1432 if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
1436 /* Retrieve the list of known public IPs from the node */
1437 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1442 &ipalloc_state->known_public_ips[j]);
1445 ("Failed to read known public IPs from node: %u\n",
1450 if (ctdb->do_checkpublicip) {
1451 verify_remote_ip_allocation(ctdb,
1452 ipalloc_state->known_public_ips[j],
1456 /* Retrieve the list of available public IPs from the node */
1457 ret = ctdb_ctrl_get_public_ips_flags(ctdb,
1461 CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
1462 &ipalloc_state->available_public_ips[j]);
1465 ("Failed to read available public IPs from node: %u\n",
1474 static struct public_ip_list *
1475 create_merged_ip_list(struct ctdb_context *ctdb, struct ipalloc_state *ipalloc_state)
1478 struct public_ip_list *ip_list;
1479 struct ctdb_public_ip_list_old *public_ips;
1481 TALLOC_FREE(ctdb->ip_tree);
1482 ctdb->ip_tree = trbt_create(ctdb, 0);
1484 for (i=0; i < ctdb->num_nodes; i++) {
1485 public_ips = ipalloc_state->known_public_ips[i];
1487 if (ctdb->nodes[i]->flags & NODE_FLAGS_DELETED) {
1491 /* there were no public ips for this node */
1492 if (public_ips == NULL) {
1496 for (j=0; j < public_ips->num; j++) {
1497 struct public_ip_list *tmp_ip;
1499 tmp_ip = talloc_zero(ctdb->ip_tree, struct public_ip_list);
1500 CTDB_NO_MEMORY_NULL(ctdb, tmp_ip);
1501 /* Do not use information about IP addresses hosted
1502 * on other nodes, it may not be accurate */
1503 if (public_ips->ips[j].pnn == ctdb->nodes[i]->pnn) {
1504 tmp_ip->pnn = public_ips->ips[j].pnn;
1508 tmp_ip->addr = public_ips->ips[j].addr;
1509 tmp_ip->next = NULL;
1511 trbt_insertarray32_callback(ctdb->ip_tree,
1512 IP_KEYLEN, ip_key(&public_ips->ips[j].addr),
1519 trbt_traversearray32(ctdb->ip_tree, IP_KEYLEN, getips_count_callback, &ip_list);
1525 * This is the length of the longtest common prefix between the IPs.
1526 * It is calculated by XOR-ing the 2 IPs together and counting the
1527 * number of leading zeroes. The implementation means that all
1528 * addresses end up being 128 bits long.
1530 * FIXME? Should we consider IPv4 and IPv6 separately given that the
1531 * 12 bytes of 0 prefix padding will hurt the algorithm if there are
1532 * lots of nodes and IP addresses?
1534 static uint32_t ip_distance(ctdb_sock_addr *ip1, ctdb_sock_addr *ip2)
1536 uint32_t ip1_k[IP_KEYLEN];
1541 uint32_t distance = 0;
1543 memcpy(ip1_k, ip_key(ip1), sizeof(ip1_k));
1545 for (i=0; i<IP_KEYLEN; i++) {
1546 x = ip1_k[i] ^ t[i];
1550 /* Count number of leading zeroes.
1551 * FIXME? This could be optimised...
1553 while ((x & (1 << 31)) == 0) {
1563 /* Calculate the IP distance for the given IP relative to IPs on the
1564 given node. The ips argument is generally the all_ips variable
1565 used in the main part of the algorithm.
1567 static uint32_t ip_distance_2_sum(ctdb_sock_addr *ip,
1568 struct public_ip_list *ips,
1571 struct public_ip_list *t;
1576 for (t=ips; t != NULL; t=t->next) {
1577 if (t->pnn != pnn) {
1581 /* Optimisation: We never calculate the distance
1582 * between an address and itself. This allows us to
1583 * calculate the effect of removing an address from a
1584 * node by simply calculating the distance between
1585 * that address and all of the exitsing addresses.
1586 * Moreover, we assume that we're only ever dealing
1587 * with addresses from all_ips so we can identify an
1588 * address via a pointer rather than doing a more
1589 * expensive address comparison. */
1590 if (&(t->addr) == ip) {
1594 d = ip_distance(ip, &(t->addr));
1595 sum += d * d; /* Cheaper than pulling in math.h :-) */
1601 /* Return the LCP2 imbalance metric for addresses currently assigned
1604 static uint32_t lcp2_imbalance(struct public_ip_list * all_ips, int pnn)
1606 struct public_ip_list *t;
1608 uint32_t imbalance = 0;
1610 for (t=all_ips; t!=NULL; t=t->next) {
1611 if (t->pnn != pnn) {
1614 /* Pass the rest of the IPs rather than the whole
1617 imbalance += ip_distance_2_sum(&(t->addr), t->next, pnn);
1623 /* Allocate any unassigned IPs just by looping through the IPs and
1624 * finding the best node for each.
1626 static void basic_allocate_unassigned(struct ipalloc_state *ipalloc_state,
1627 struct public_ip_list *all_ips)
1629 struct public_ip_list *tmp_ip;
1631 /* loop over all ip's and find a physical node to cover for
1634 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1635 if (tmp_ip->pnn == -1) {
1636 if (find_takeover_node(ipalloc_state,
1638 DEBUG(DEBUG_WARNING,
1639 ("Failed to find node to cover ip %s\n",
1640 ctdb_addr_to_str(&tmp_ip->addr)));
1646 /* Basic non-deterministic rebalancing algorithm.
1648 static void basic_failback(struct ipalloc_state *ipalloc_state,
1649 struct public_ip_list *all_ips,
1653 int maxnode, maxnum, minnode, minnum, num, retries;
1654 struct public_ip_list *tmp_ip;
1656 numnodes = ipalloc_state->num;
1663 /* for each ip address, loop over all nodes that can serve
1664 this ip and make sure that the difference between the node
1665 serving the most and the node serving the least ip's are
1668 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1669 if (tmp_ip->pnn == -1) {
1673 /* Get the highest and lowest number of ips's served by any
1674 valid node which can serve this ip.
1678 for (i=0; i<numnodes; i++) {
1679 /* only check nodes that can actually serve this ip */
1680 if (!can_node_takeover_ip(ipalloc_state, i,
1682 /* no it couldnt so skip to the next node */
1686 num = node_ip_coverage(i, all_ips);
1687 if (maxnode == -1) {
1696 if (minnode == -1) {
1706 if (maxnode == -1) {
1707 DEBUG(DEBUG_WARNING,(__location__ " Could not find maxnode. May not be able to serve ip '%s'\n",
1708 ctdb_addr_to_str(&tmp_ip->addr)));
1713 /* if the spread between the smallest and largest coverage by
1714 a node is >=2 we steal one of the ips from the node with
1715 most coverage to even things out a bit.
1716 try to do this a limited number of times since we dont
1717 want to spend too much time balancing the ip coverage.
1719 if ( (maxnum > minnum+1)
1720 && (retries < (num_ips + 5)) ){
1721 struct public_ip_list *tmp;
1723 /* Reassign one of maxnode's VNNs */
1724 for (tmp=all_ips;tmp;tmp=tmp->next) {
1725 if (tmp->pnn == maxnode) {
1726 (void)find_takeover_node(ipalloc_state,
1737 static bool lcp2_init(struct ipalloc_state *ipalloc_state,
1738 struct public_ip_list *all_ips,
1739 uint32_t *force_rebalance_nodes,
1740 uint32_t **lcp2_imbalances,
1741 bool **rebalance_candidates)
1744 struct public_ip_list *tmp_ip;
1746 numnodes = ipalloc_state->num;
1748 *rebalance_candidates = talloc_array(ipalloc_state, bool, numnodes);
1749 if (*rebalance_candidates == NULL) {
1750 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1753 *lcp2_imbalances = talloc_array(ipalloc_state, uint32_t, numnodes);
1754 if (*lcp2_imbalances == NULL) {
1755 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1759 for (i=0; i<numnodes; i++) {
1760 (*lcp2_imbalances)[i] = lcp2_imbalance(all_ips, i);
1761 /* First step: assume all nodes are candidates */
1762 (*rebalance_candidates)[i] = true;
1765 /* 2nd step: if a node has IPs assigned then it must have been
1766 * healthy before, so we remove it from consideration. This
1767 * is overkill but is all we have because we don't maintain
1768 * state between takeover runs. An alternative would be to
1769 * keep state and invalidate it every time the recovery master
1772 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1773 if (tmp_ip->pnn != -1) {
1774 (*rebalance_candidates)[tmp_ip->pnn] = false;
1778 /* 3rd step: if a node is forced to re-balance then
1779 we allow failback onto the node */
1780 if (force_rebalance_nodes == NULL) {
1783 for (i = 0; i < talloc_array_length(force_rebalance_nodes); i++) {
1784 uint32_t pnn = force_rebalance_nodes[i];
1785 if (pnn >= numnodes) {
1787 (__location__ "unknown node %u\n", pnn));
1792 ("Forcing rebalancing of IPs to node %u\n", pnn));
1793 (*rebalance_candidates)[pnn] = true;
1799 /* Allocate any unassigned addresses using the LCP2 algorithm to find
1800 * the IP/node combination that will cost the least.
1802 static void lcp2_allocate_unassigned(struct ipalloc_state *ipalloc_state,
1803 struct public_ip_list *all_ips,
1804 uint32_t *lcp2_imbalances)
1806 struct public_ip_list *tmp_ip;
1807 int dstnode, numnodes;
1810 uint32_t mindsum, dstdsum, dstimbl, minimbl;
1811 struct public_ip_list *minip;
1813 bool should_loop = true;
1814 bool have_unassigned = true;
1816 numnodes = ipalloc_state->num;
1818 while (have_unassigned && should_loop) {
1819 should_loop = false;
1821 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
1822 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES (UNASSIGNED)\n"));
1828 /* loop over each unassigned ip. */
1829 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1830 if (tmp_ip->pnn != -1) {
1834 for (dstnode=0; dstnode<numnodes; dstnode++) {
1835 /* only check nodes that can actually takeover this ip */
1836 if (!can_node_takeover_ip(ipalloc_state,
1839 /* no it couldnt so skip to the next node */
1843 dstdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, dstnode);
1844 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
1845 DEBUG(DEBUG_DEBUG,(" %s -> %d [+%d]\n",
1846 ctdb_addr_to_str(&(tmp_ip->addr)),
1848 dstimbl - lcp2_imbalances[dstnode]));
1851 if ((minnode == -1) || (dstdsum < mindsum)) {
1861 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
1863 /* If we found one then assign it to the given node. */
1864 if (minnode != -1) {
1865 minip->pnn = minnode;
1866 lcp2_imbalances[minnode] = minimbl;
1867 DEBUG(DEBUG_INFO,(" %s -> %d [+%d]\n",
1868 ctdb_addr_to_str(&(minip->addr)),
1873 /* There might be a better way but at least this is clear. */
1874 have_unassigned = false;
1875 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1876 if (tmp_ip->pnn == -1) {
1877 have_unassigned = true;
1882 /* We know if we have an unassigned addresses so we might as
1885 if (have_unassigned) {
1886 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
1887 if (tmp_ip->pnn == -1) {
1888 DEBUG(DEBUG_WARNING,("Failed to find node to cover ip %s\n",
1889 ctdb_addr_to_str(&tmp_ip->addr)));
1895 /* LCP2 algorithm for rebalancing the cluster. Given a candidate node
1896 * to move IPs from, determines the best IP/destination node
1897 * combination to move from the source node.
1899 static bool lcp2_failback_candidate(struct ipalloc_state *ipalloc_state,
1900 struct public_ip_list *all_ips,
1902 uint32_t *lcp2_imbalances,
1903 bool *rebalance_candidates)
1905 int dstnode, mindstnode, numnodes;
1906 uint32_t srcimbl, srcdsum, dstimbl, dstdsum;
1907 uint32_t minsrcimbl, mindstimbl;
1908 struct public_ip_list *minip;
1909 struct public_ip_list *tmp_ip;
1911 /* Find an IP and destination node that best reduces imbalance. */
1918 numnodes = ipalloc_state->num;
1920 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
1921 DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES FROM %d [%d]\n",
1922 srcnode, lcp2_imbalances[srcnode]));
1924 for (tmp_ip=all_ips; tmp_ip; tmp_ip=tmp_ip->next) {
1925 /* Only consider addresses on srcnode. */
1926 if (tmp_ip->pnn != srcnode) {
1930 /* What is this IP address costing the source node? */
1931 srcdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, srcnode);
1932 srcimbl = lcp2_imbalances[srcnode] - srcdsum;
1934 /* Consider this IP address would cost each potential
1935 * destination node. Destination nodes are limited to
1936 * those that are newly healthy, since we don't want
1937 * to do gratuitous failover of IPs just to make minor
1938 * balance improvements.
1940 for (dstnode=0; dstnode<numnodes; dstnode++) {
1941 if (!rebalance_candidates[dstnode]) {
1945 /* only check nodes that can actually takeover this ip */
1946 if (!can_node_takeover_ip(ipalloc_state, dstnode,
1948 /* no it couldnt so skip to the next node */
1952 dstdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, dstnode);
1953 dstimbl = lcp2_imbalances[dstnode] + dstdsum;
1954 DEBUG(DEBUG_DEBUG,(" %d [%d] -> %s -> %d [+%d]\n",
1956 ctdb_addr_to_str(&(tmp_ip->addr)),
1959 if ((dstimbl < lcp2_imbalances[srcnode]) &&
1960 (dstdsum < srcdsum) && \
1961 ((mindstnode == -1) || \
1962 ((srcimbl + dstimbl) < (minsrcimbl + mindstimbl)))) {
1965 minsrcimbl = srcimbl;
1966 mindstnode = dstnode;
1967 mindstimbl = dstimbl;
1971 DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
1973 if (mindstnode != -1) {
1974 /* We found a move that makes things better... */
1975 DEBUG(DEBUG_INFO,("%d [%d] -> %s -> %d [+%d]\n",
1976 srcnode, minsrcimbl - lcp2_imbalances[srcnode],
1977 ctdb_addr_to_str(&(minip->addr)),
1978 mindstnode, mindstimbl - lcp2_imbalances[mindstnode]));
1981 lcp2_imbalances[srcnode] = minsrcimbl;
1982 lcp2_imbalances[mindstnode] = mindstimbl;
1983 minip->pnn = mindstnode;
1992 struct lcp2_imbalance_pnn {
1997 static int lcp2_cmp_imbalance_pnn(const void * a, const void * b)
1999 const struct lcp2_imbalance_pnn * lipa = (const struct lcp2_imbalance_pnn *) a;
2000 const struct lcp2_imbalance_pnn * lipb = (const struct lcp2_imbalance_pnn *) b;
2002 if (lipa->imbalance > lipb->imbalance) {
2004 } else if (lipa->imbalance == lipb->imbalance) {
2011 /* LCP2 algorithm for rebalancing the cluster. This finds the source
2012 * node with the highest LCP2 imbalance, and then determines the best
2013 * IP/destination node combination to move from the source node.
2015 static void lcp2_failback(struct ipalloc_state *ipalloc_state,
2016 struct public_ip_list *all_ips,
2017 uint32_t *lcp2_imbalances,
2018 bool *rebalance_candidates)
2021 struct lcp2_imbalance_pnn * lips;
2024 numnodes = ipalloc_state->num;
2027 /* Put the imbalances and nodes into an array, sort them and
2028 * iterate through candidates. Usually the 1st one will be
2029 * used, so this doesn't cost much...
2031 DEBUG(DEBUG_DEBUG,("+++++++++++++++++++++++++++++++++++++++++\n"));
2032 DEBUG(DEBUG_DEBUG,("Selecting most imbalanced node from:\n"));
2033 lips = talloc_array(ipalloc_state, struct lcp2_imbalance_pnn, numnodes);
2034 for (i=0; i<numnodes; i++) {
2035 lips[i].imbalance = lcp2_imbalances[i];
2037 DEBUG(DEBUG_DEBUG,(" %d [%d]\n", i, lcp2_imbalances[i]));
2039 qsort(lips, numnodes, sizeof(struct lcp2_imbalance_pnn),
2040 lcp2_cmp_imbalance_pnn);
2043 for (i=0; i<numnodes; i++) {
2044 /* This means that all nodes had 0 or 1 addresses, so
2045 * can't be imbalanced.
2047 if (lips[i].imbalance == 0) {
2051 if (lcp2_failback_candidate(ipalloc_state,
2055 rebalance_candidates)) {
2067 static void unassign_unsuitable_ips(struct ipalloc_state *ipalloc_state,
2068 struct public_ip_list *all_ips)
2070 struct public_ip_list *tmp_ip;
2072 /* verify that the assigned nodes can serve that public ip
2073 and set it to -1 if not
2075 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
2076 if (tmp_ip->pnn == -1) {
2079 if (!can_node_host_ip(ipalloc_state, tmp_ip->pnn,
2081 /* this node can not serve this ip. */
2082 DEBUG(DEBUG_DEBUG,("Unassign IP: %s from %d\n",
2083 ctdb_addr_to_str(&(tmp_ip->addr)),
2090 static bool ip_alloc_deterministic_ips(struct ipalloc_state *ipalloc_state,
2091 struct public_ip_list *all_ips)
2093 struct public_ip_list *tmp_ip;
2096 numnodes = ipalloc_state->num;
2098 DEBUG(DEBUG_NOTICE,("Deterministic IPs enabled. Resetting all ip allocations\n"));
2099 /* Allocate IPs to nodes in a modulo fashion so that IPs will
2100 * always be allocated the same way for a specific set of
2101 * available/unavailable nodes.
2104 for (i=0,tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next,i++) {
2105 tmp_ip->pnn = i % numnodes;
2108 /* IP failback doesn't make sense with deterministic
2109 * IPs, since the modulo step above implicitly fails
2110 * back IPs to their "home" node.
2112 if (1 == ipalloc_state->no_ip_failback) {
2113 DEBUG(DEBUG_WARNING, ("WARNING: 'NoIPFailback' set but ignored - incompatible with 'DeterministicIPs\n"));
2116 unassign_unsuitable_ips(ipalloc_state, all_ips);
2118 basic_allocate_unassigned(ipalloc_state, all_ips);
2120 /* No failback here! */
2125 static bool ip_alloc_nondeterministic_ips(struct ipalloc_state *ipalloc_state,
2126 struct public_ip_list *all_ips)
2128 /* This should be pushed down into basic_failback. */
2129 struct public_ip_list *tmp_ip;
2131 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
2135 unassign_unsuitable_ips(ipalloc_state, all_ips);
2137 basic_allocate_unassigned(ipalloc_state, all_ips);
2139 /* If we don't want IPs to fail back then don't rebalance IPs. */
2140 if (1 == ipalloc_state->no_ip_failback) {
2144 /* Now, try to make sure the ip adresses are evenly distributed
2147 basic_failback(ipalloc_state, all_ips, num_ips);
2152 static bool ip_alloc_lcp2(struct ipalloc_state *ipalloc_state,
2153 struct public_ip_list *all_ips,
2154 uint32_t *force_rebalance_nodes)
2156 uint32_t *lcp2_imbalances;
2157 bool *rebalance_candidates;
2158 int numnodes, num_rebalance_candidates, i;
2161 unassign_unsuitable_ips(ipalloc_state, all_ips);
2163 if (!lcp2_init(ipalloc_state, all_ips,force_rebalance_nodes,
2164 &lcp2_imbalances, &rebalance_candidates)) {
2169 lcp2_allocate_unassigned(ipalloc_state, all_ips, lcp2_imbalances);
2171 /* If we don't want IPs to fail back then don't rebalance IPs. */
2172 if (1 == ipalloc_state->no_ip_failback) {
2176 /* It is only worth continuing if we have suitable target
2177 * nodes to transfer IPs to. This check is much cheaper than
2180 numnodes = ipalloc_state->num;
2181 num_rebalance_candidates = 0;
2182 for (i=0; i<numnodes; i++) {
2183 if (rebalance_candidates[i]) {
2184 num_rebalance_candidates++;
2187 if (num_rebalance_candidates == 0) {
2191 /* Now, try to make sure the ip adresses are evenly distributed
2194 lcp2_failback(ipalloc_state, all_ips,
2195 lcp2_imbalances, rebalance_candidates);
2201 static bool all_nodes_are_disabled(struct ctdb_node_map_old *nodemap)
2205 for (i=0;i<nodemap->num;i++) {
2206 if (!(nodemap->nodes[i].flags & (NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED))) {
2207 /* Found one completely healthy node */
2215 /* The calculation part of the IP allocation algorithm. */
2216 static bool ctdb_takeover_run_core(struct ipalloc_state *ipalloc_state,
2217 struct public_ip_list *all_ips,
2218 uint32_t *force_rebalance_nodes)
2222 switch (ipalloc_state->algorithm) {
2224 ret = ip_alloc_lcp2(ipalloc_state, all_ips,
2225 force_rebalance_nodes);
2227 case IPALLOC_DETERMINISTIC:
2228 ret = ip_alloc_deterministic_ips(ipalloc_state, all_ips);
2230 case IPALLOC_NONDETERMINISTIC:
2231 ret = ip_alloc_nondeterministic_ips(ipalloc_state, all_ips);
2235 /* at this point ->pnn is the node which will own each IP
2236 or -1 if there is no node that can cover this ip
2242 struct get_tunable_callback_data {
2243 const char *tunable;
2248 static void get_tunable_callback(struct ctdb_context *ctdb, uint32_t pnn,
2249 int32_t res, TDB_DATA outdata,
2252 struct get_tunable_callback_data *cd =
2253 (struct get_tunable_callback_data *)callback;
2257 /* Already handled in fail callback */
2261 if (outdata.dsize != sizeof(uint32_t)) {
2262 DEBUG(DEBUG_ERR,("Wrong size of returned data when reading \"%s\" tunable from node %d. Expected %d bytes but received %d bytes\n",
2263 cd->tunable, pnn, (int)sizeof(uint32_t),
2264 (int)outdata.dsize));
2269 size = talloc_array_length(cd->out);
2271 DEBUG(DEBUG_ERR,("Got %s reply from node %d but nodemap only has %d entries\n",
2272 cd->tunable, pnn, size));
2277 cd->out[pnn] = *(uint32_t *)outdata.dptr;
2280 static void get_tunable_fail_callback(struct ctdb_context *ctdb, uint32_t pnn,
2281 int32_t res, TDB_DATA outdata,
2284 struct get_tunable_callback_data *cd =
2285 (struct get_tunable_callback_data *)callback;
2290 ("Timed out getting tunable \"%s\" from node %d\n",
2296 DEBUG(DEBUG_WARNING,
2297 ("Tunable \"%s\" not implemented on node %d\n",
2302 ("Unexpected error getting tunable \"%s\" from node %d\n",
2308 static uint32_t *get_tunable_from_nodes(struct ctdb_context *ctdb,
2309 TALLOC_CTX *tmp_ctx,
2310 struct ctdb_node_map_old *nodemap,
2311 const char *tunable,
2312 uint32_t default_value)
2315 struct ctdb_control_get_tunable *t;
2318 struct get_tunable_callback_data callback_data;
2321 tvals = talloc_array(tmp_ctx, uint32_t, nodemap->num);
2322 CTDB_NO_MEMORY_NULL(ctdb, tvals);
2323 for (i=0; i<nodemap->num; i++) {
2324 tvals[i] = default_value;
2327 callback_data.out = tvals;
2328 callback_data.tunable = tunable;
2329 callback_data.fatal = false;
2331 data.dsize = offsetof(struct ctdb_control_get_tunable, name) + strlen(tunable) + 1;
2332 data.dptr = talloc_size(tmp_ctx, data.dsize);
2333 t = (struct ctdb_control_get_tunable *)data.dptr;
2334 t->length = strlen(tunable)+1;
2335 memcpy(t->name, tunable, t->length);
2336 nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2337 if (ctdb_client_async_control(ctdb, CTDB_CONTROL_GET_TUNABLE,
2338 nodes, 0, TAKEOVER_TIMEOUT(),
2340 get_tunable_callback,
2341 get_tunable_fail_callback,
2342 &callback_data) != 0) {
2343 if (callback_data.fatal) {
2349 talloc_free(data.dptr);
2354 /* Set internal flags for IP allocation:
2356 * Set NOIPTAKOVER ip flags from per-node NoIPTakeover tunable
2357 * Set NOIPHOST ip flag for each INACTIVE node
2358 * if all nodes are disabled:
2359 * Set NOIPHOST ip flags from per-node NoIPHostOnAllDisabled tunable
2361 * Set NOIPHOST ip flags for disabled nodes
2363 static void set_ipflags_internal(struct ipalloc_state *ipalloc_state,
2364 struct ctdb_node_map_old *nodemap,
2365 uint32_t *tval_noiptakeover,
2366 uint32_t *tval_noiphostonalldisabled)
2370 /* IP flags cleared at this point - implicit due to talloc_zero */
2372 for (i=0;i<nodemap->num;i++) {
2373 /* Can not take IPs on node with NoIPTakeover set */
2374 if (tval_noiptakeover[i] != 0) {
2375 ipalloc_state->ipflags[i].noiptakeover = true;
2378 /* Can not host IPs on INACTIVE node */
2379 if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
2380 ipalloc_state->ipflags[i].noiphost = true;
2384 if (all_nodes_are_disabled(nodemap)) {
2385 /* If all nodes are disabled, can not host IPs on node
2386 * with NoIPHostOnAllDisabled set
2388 for (i=0;i<nodemap->num;i++) {
2389 if (tval_noiphostonalldisabled[i] != 0) {
2390 ipalloc_state->ipflags[i].noiphost = true;
2394 /* If some nodes are not disabled, then can not host
2395 * IPs on DISABLED node
2397 for (i=0;i<nodemap->num;i++) {
2398 if (nodemap->nodes[i].flags & NODE_FLAGS_DISABLED) {
2399 ipalloc_state->ipflags[i].noiphost = true;
2405 static bool set_ipflags(struct ctdb_context *ctdb,
2406 struct ipalloc_state *ipalloc_state,
2407 struct ctdb_node_map_old *nodemap)
2409 uint32_t *tval_noiptakeover;
2410 uint32_t *tval_noiphostonalldisabled;
2412 tval_noiptakeover = get_tunable_from_nodes(ctdb, ipalloc_state, nodemap,
2414 if (tval_noiptakeover == NULL) {
2418 tval_noiphostonalldisabled =
2419 get_tunable_from_nodes(ctdb, ipalloc_state, nodemap,
2420 "NoIPHostOnAllDisabled", 0);
2421 if (tval_noiphostonalldisabled == NULL) {
2422 /* Caller frees tmp_ctx */
2426 set_ipflags_internal(ipalloc_state, nodemap,
2428 tval_noiphostonalldisabled);
2430 talloc_free(tval_noiptakeover);
2431 talloc_free(tval_noiphostonalldisabled);
2436 static struct ipalloc_state * ipalloc_state_init(struct ctdb_context *ctdb,
2437 TALLOC_CTX *mem_ctx)
2439 struct ipalloc_state *ipalloc_state =
2440 talloc_zero(mem_ctx, struct ipalloc_state);
2441 if (ipalloc_state == NULL) {
2442 DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
2446 ipalloc_state->num = ctdb->num_nodes;
2447 ipalloc_state->known_public_ips =
2448 talloc_zero_array(ipalloc_state,
2449 struct ctdb_public_ip_list_old *,
2450 ipalloc_state->num);
2451 if (ipalloc_state->known_public_ips == NULL) {
2452 DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
2453 talloc_free(ipalloc_state);
2456 ipalloc_state->available_public_ips =
2457 talloc_zero_array(ipalloc_state,
2458 struct ctdb_public_ip_list_old *,
2459 ipalloc_state->num);
2460 if (ipalloc_state->available_public_ips == NULL) {
2461 DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
2462 talloc_free(ipalloc_state);
2465 ipalloc_state->ipflags =
2466 talloc_zero_array(ipalloc_state,
2467 struct ctdb_ipflags,
2468 ipalloc_state->num);
2469 if (ipalloc_state->ipflags == NULL) {
2470 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
2471 talloc_free(ipalloc_state);
2475 if (1 == ctdb->tunable.lcp2_public_ip_assignment) {
2476 ipalloc_state->algorithm = IPALLOC_LCP2;
2477 } else if (1 == ctdb->tunable.deterministic_public_ips) {
2478 ipalloc_state->algorithm = IPALLOC_DETERMINISTIC;
2480 ipalloc_state->algorithm = IPALLOC_NONDETERMINISTIC;
2483 ipalloc_state->no_ip_failback = ctdb->tunable.no_ip_failback;
2485 return ipalloc_state;
2488 struct iprealloc_callback_data {
2491 client_async_callback fail_callback;
2492 void *fail_callback_data;
2493 struct ctdb_node_map_old *nodemap;
2496 static void iprealloc_fail_callback(struct ctdb_context *ctdb, uint32_t pnn,
2497 int32_t res, TDB_DATA outdata,
2501 struct iprealloc_callback_data *cd =
2502 (struct iprealloc_callback_data *)callback;
2504 numnodes = talloc_array_length(cd->retry_nodes);
2505 if (pnn > numnodes) {
2507 ("ipreallocated failure from node %d, "
2508 "but only %d nodes in nodemap\n",
2513 /* Can't run the "ipreallocated" event on a INACTIVE node */
2514 if (cd->nodemap->nodes[pnn].flags & NODE_FLAGS_INACTIVE) {
2515 DEBUG(DEBUG_WARNING,
2516 ("ipreallocated failed on inactive node %d, ignoring\n",
2523 /* If the control timed out then that's a real error,
2524 * so call the real fail callback
2526 if (cd->fail_callback) {
2527 cd->fail_callback(ctdb, pnn, res, outdata,
2528 cd->fail_callback_data);
2530 DEBUG(DEBUG_WARNING,
2531 ("iprealloc timed out but no callback registered\n"));
2535 /* If not a timeout then either the ipreallocated
2536 * eventscript (or some setup) failed. This might
2537 * have failed because the IPREALLOCATED control isn't
2538 * implemented - right now there is no way of knowing
2539 * because the error codes are all folded down to -1.
2540 * Consider retrying using EVENTSCRIPT control...
2542 DEBUG(DEBUG_WARNING,
2543 ("ipreallocated failure from node %d, flagging retry\n",
2545 cd->retry_nodes[pnn] = true;
2550 struct takeover_callback_data {
2552 client_async_callback fail_callback;
2553 void *fail_callback_data;
2554 struct ctdb_node_map_old *nodemap;
2557 static void takeover_run_fail_callback(struct ctdb_context *ctdb,
2558 uint32_t node_pnn, int32_t res,
2559 TDB_DATA outdata, void *callback_data)
2561 struct takeover_callback_data *cd =
2562 talloc_get_type_abort(callback_data,
2563 struct takeover_callback_data);
2566 for (i = 0; i < cd->nodemap->num; i++) {
2567 if (node_pnn == cd->nodemap->nodes[i].pnn) {
2572 if (i == cd->nodemap->num) {
2573 DEBUG(DEBUG_ERR, (__location__ " invalid PNN %u\n", node_pnn));
2577 if (!cd->node_failed[i]) {
2578 cd->node_failed[i] = true;
2579 cd->fail_callback(ctdb, node_pnn, res, outdata,
2580 cd->fail_callback_data);
2585 make any IP alias changes for public addresses that are necessary
2587 int ctdb_takeover_run(struct ctdb_context *ctdb, struct ctdb_node_map_old *nodemap,
2588 uint32_t *force_rebalance_nodes,
2589 client_async_callback fail_callback, void *callback_data)
2592 struct ctdb_public_ip ip;
2594 struct public_ip_list *all_ips, *tmp_ip;
2596 struct timeval timeout;
2597 struct client_async_data *async_data;
2598 struct ctdb_client_control_state *state;
2599 TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
2600 struct ipalloc_state *ipalloc_state;
2601 struct takeover_callback_data *takeover_data;
2602 struct iprealloc_callback_data iprealloc_data;
2607 * ip failover is completely disabled, just send out the
2608 * ipreallocated event.
2610 if (ctdb->tunable.disable_ip_failover != 0) {
2614 ipalloc_state = ipalloc_state_init(ctdb, tmp_ctx);
2615 if (ipalloc_state == NULL) {
2616 talloc_free(tmp_ctx);
2620 if (!set_ipflags(ctdb, ipalloc_state, nodemap)) {
2621 DEBUG(DEBUG_ERR,("Failed to set IP flags - aborting takeover run\n"));
2622 talloc_free(tmp_ctx);
2626 /* Fetch known/available public IPs from each active node */
2627 ret = ctdb_reload_remote_public_ips(ctdb, ipalloc_state, nodemap);
2629 talloc_free(tmp_ctx);
2633 /* Short-circuit IP allocation if no node has available IPs */
2634 can_host_ips = false;
2635 for (i=0; i < ipalloc_state->num; i++) {
2636 if (ipalloc_state->available_public_ips[i] != NULL) {
2637 can_host_ips = true;
2640 if (!can_host_ips) {
2641 DEBUG(DEBUG_WARNING,("No nodes available to host public IPs yet\n"));
2645 /* since nodes only know about those public addresses that
2646 can be served by that particular node, no single node has
2647 a full list of all public addresses that exist in the cluster.
2648 Walk over all node structures and create a merged list of
2649 all public addresses that exist in the cluster.
2651 keep the tree of ips around as ctdb->ip_tree
2653 all_ips = create_merged_ip_list(ctdb, ipalloc_state);
2655 /* Do the IP reassignment calculations */
2656 ctdb_takeover_run_core(ipalloc_state,
2657 all_ips, force_rebalance_nodes);
2659 /* Now tell all nodes to release any public IPs should not
2660 * host. This will be a NOOP on nodes that don't currently
2661 * hold the given IP.
2663 takeover_data = talloc_zero(tmp_ctx, struct takeover_callback_data);
2664 CTDB_NO_MEMORY_FATAL(ctdb, takeover_data);
2666 takeover_data->node_failed = talloc_zero_array(tmp_ctx,
2667 bool, nodemap->num);
2668 CTDB_NO_MEMORY_FATAL(ctdb, takeover_data->node_failed);
2669 takeover_data->fail_callback = fail_callback;
2670 takeover_data->fail_callback_data = callback_data;
2671 takeover_data->nodemap = nodemap;
2673 async_data = talloc_zero(tmp_ctx, struct client_async_data);
2674 CTDB_NO_MEMORY_FATAL(ctdb, async_data);
2676 async_data->fail_callback = takeover_run_fail_callback;
2677 async_data->callback_data = takeover_data;
2679 ZERO_STRUCT(ip); /* Avoid valgrind warnings for union */
2681 /* Send a RELEASE_IP to all nodes that should not be hosting
2682 * each IP. For each IP, all but one of these will be
2683 * redundant. However, the redundant ones are used to tell
2684 * nodes which node should be hosting the IP so that commands
2685 * like "ctdb ip" can display a particular nodes idea of who
2686 * is hosting what. */
2687 for (i=0;i<nodemap->num;i++) {
2688 /* don't talk to unconnected nodes, but do talk to banned nodes */
2689 if (nodemap->nodes[i].flags & NODE_FLAGS_DISCONNECTED) {
2693 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
2694 if (tmp_ip->pnn == nodemap->nodes[i].pnn) {
2695 /* This node should be serving this
2696 vnn so don't tell it to release the ip
2700 ip.pnn = tmp_ip->pnn;
2701 ip.addr = tmp_ip->addr;
2703 timeout = TAKEOVER_TIMEOUT();
2704 data.dsize = sizeof(ip);
2705 data.dptr = (uint8_t *)&ip;
2706 state = ctdb_control_send(ctdb, nodemap->nodes[i].pnn,
2707 0, CTDB_CONTROL_RELEASE_IP, 0,
2710 if (state == NULL) {
2711 DEBUG(DEBUG_ERR,(__location__ " Failed to call async control CTDB_CONTROL_RELEASE_IP to node %u\n", nodemap->nodes[i].pnn));
2712 talloc_free(tmp_ctx);
2716 ctdb_client_async_add(async_data, state);
2719 if (ctdb_client_async_wait(ctdb, async_data) != 0) {
2720 DEBUG(DEBUG_ERR,(__location__ " Async control CTDB_CONTROL_RELEASE_IP failed\n"));
2721 talloc_free(tmp_ctx);
2724 talloc_free(async_data);
2727 /* For each IP, send a TAKOVER_IP to the node that should be
2728 * hosting it. Many of these will often be redundant (since
2729 * the allocation won't have changed) but they can be useful
2730 * to recover from inconsistencies. */
2731 async_data = talloc_zero(tmp_ctx, struct client_async_data);
2732 CTDB_NO_MEMORY_FATAL(ctdb, async_data);
2734 async_data->fail_callback = fail_callback;
2735 async_data->callback_data = callback_data;
2737 for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
2738 if (tmp_ip->pnn == -1) {
2739 /* this IP won't be taken over */
2743 ip.pnn = tmp_ip->pnn;
2744 ip.addr = tmp_ip->addr;
2746 timeout = TAKEOVER_TIMEOUT();
2747 data.dsize = sizeof(ip);
2748 data.dptr = (uint8_t *)&ip;
2749 state = ctdb_control_send(ctdb, tmp_ip->pnn,
2750 0, CTDB_CONTROL_TAKEOVER_IP, 0,
2751 data, async_data, &timeout, NULL);
2752 if (state == NULL) {
2753 DEBUG(DEBUG_ERR,(__location__ " Failed to call async control CTDB_CONTROL_TAKEOVER_IP to node %u\n", tmp_ip->pnn));
2754 talloc_free(tmp_ctx);
2758 ctdb_client_async_add(async_data, state);
2760 if (ctdb_client_async_wait(ctdb, async_data) != 0) {
2761 DEBUG(DEBUG_ERR,(__location__ " Async control CTDB_CONTROL_TAKEOVER_IP failed\n"));
2762 talloc_free(tmp_ctx);
2768 * Tell all nodes to run eventscripts to process the
2769 * "ipreallocated" event. This can do a lot of things,
2770 * including restarting services to reconfigure them if public
2771 * IPs have moved. Once upon a time this event only used to
2774 retry_data = talloc_zero_array(tmp_ctx, bool, nodemap->num);
2775 CTDB_NO_MEMORY_FATAL(ctdb, retry_data);
2776 iprealloc_data.retry_nodes = retry_data;
2777 iprealloc_data.retry_count = 0;
2778 iprealloc_data.fail_callback = fail_callback;
2779 iprealloc_data.fail_callback_data = callback_data;
2780 iprealloc_data.nodemap = nodemap;
2782 nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
2783 ret = ctdb_client_async_control(ctdb, CTDB_CONTROL_IPREALLOCATED,
2784 nodes, 0, TAKEOVER_TIMEOUT(),
2786 NULL, iprealloc_fail_callback,
2789 /* If the control failed then we should retry to any
2790 * nodes flagged by iprealloc_fail_callback using the
2791 * EVENTSCRIPT control. This is a best-effort at
2792 * backward compatiblity when running a mixed cluster
2793 * where some nodes have not yet been upgraded to
2794 * support the IPREALLOCATED control.
2796 DEBUG(DEBUG_WARNING,
2797 ("Retry ipreallocated to some nodes using eventscript control\n"));
2799 nodes = talloc_array(tmp_ctx, uint32_t,
2800 iprealloc_data.retry_count);
2801 CTDB_NO_MEMORY_FATAL(ctdb, nodes);
2804 for (i=0; i<nodemap->num; i++) {
2805 if (iprealloc_data.retry_nodes[i]) {
2811 data.dptr = discard_const("ipreallocated");
2812 data.dsize = strlen((char *)data.dptr) + 1;
2813 ret = ctdb_client_async_control(ctdb,
2814 CTDB_CONTROL_RUN_EVENTSCRIPTS,
2815 nodes, 0, TAKEOVER_TIMEOUT(),
2817 NULL, fail_callback,
2820 DEBUG(DEBUG_ERR, (__location__ " failed to send control to run eventscripts with \"ipreallocated\"\n"));
2824 talloc_free(tmp_ctx);
2830 destroy a ctdb_client_ip structure
2832 static int ctdb_client_ip_destructor(struct ctdb_client_ip *ip)
2834 DEBUG(DEBUG_DEBUG,("destroying client tcp for %s:%u (client_id %u)\n",
2835 ctdb_addr_to_str(&ip->addr),
2836 ntohs(ip->addr.ip.sin_port),
2839 DLIST_REMOVE(ip->ctdb->client_ip_list, ip);
2844 called by a client to inform us of a TCP connection that it is managing
2845 that should tickled with an ACK when IP takeover is done
2847 int32_t ctdb_control_tcp_client(struct ctdb_context *ctdb, uint32_t client_id,
2850 struct ctdb_client *client = reqid_find(ctdb->idr, client_id, struct ctdb_client);
2851 struct ctdb_connection *tcp_sock = NULL;
2852 struct ctdb_tcp_list *tcp;
2853 struct ctdb_connection t;
2856 struct ctdb_client_ip *ip;
2857 struct ctdb_vnn *vnn;
2858 ctdb_sock_addr addr;
2860 /* If we don't have public IPs, tickles are useless */
2861 if (ctdb->vnn == NULL) {
2865 tcp_sock = (struct ctdb_connection *)indata.dptr;
2867 addr = tcp_sock->src;
2868 ctdb_canonicalize_ip(&addr, &tcp_sock->src);
2869 addr = tcp_sock->dst;
2870 ctdb_canonicalize_ip(&addr, &tcp_sock->dst);
2873 memcpy(&addr, &tcp_sock->dst, sizeof(addr));
2874 vnn = find_public_ip_vnn(ctdb, &addr);
2876 switch (addr.sa.sa_family) {
2878 if (ntohl(addr.ip.sin_addr.s_addr) != INADDR_LOOPBACK) {
2879 DEBUG(DEBUG_ERR,("Could not add client IP %s. This is not a public address.\n",
2880 ctdb_addr_to_str(&addr)));
2884 DEBUG(DEBUG_ERR,("Could not add client IP %s. This is not a public ipv6 address.\n",
2885 ctdb_addr_to_str(&addr)));
2888 DEBUG(DEBUG_ERR,(__location__ " Unknown family type %d\n", addr.sa.sa_family));
2894 if (vnn->pnn != ctdb->pnn) {
2895 DEBUG(DEBUG_ERR,("Attempt to register tcp client for IP %s we don't hold - failing (client_id %u pid %u)\n",
2896 ctdb_addr_to_str(&addr),
2897 client_id, client->pid));
2898 /* failing this call will tell smbd to die */
2902 ip = talloc(client, struct ctdb_client_ip);
2903 CTDB_NO_MEMORY(ctdb, ip);
2907 ip->client_id = client_id;
2908 talloc_set_destructor(ip, ctdb_client_ip_destructor);
2909 DLIST_ADD(ctdb->client_ip_list, ip);
2911 tcp = talloc(client, struct ctdb_tcp_list);
2912 CTDB_NO_MEMORY(ctdb, tcp);
2914 tcp->connection.src = tcp_sock->src;
2915 tcp->connection.dst = tcp_sock->dst;
2917 DLIST_ADD(client->tcp_list, tcp);
2919 t.src = tcp_sock->src;
2920 t.dst = tcp_sock->dst;
2922 data.dptr = (uint8_t *)&t;
2923 data.dsize = sizeof(t);
2925 switch (addr.sa.sa_family) {
2927 DEBUG(DEBUG_INFO,("registered tcp client for %u->%s:%u (client_id %u pid %u)\n",
2928 (unsigned)ntohs(tcp_sock->dst.ip.sin_port),
2929 ctdb_addr_to_str(&tcp_sock->src),
2930 (unsigned)ntohs(tcp_sock->src.ip.sin_port), client_id, client->pid));
2933 DEBUG(DEBUG_INFO,("registered tcp client for %u->%s:%u (client_id %u pid %u)\n",
2934 (unsigned)ntohs(tcp_sock->dst.ip6.sin6_port),
2935 ctdb_addr_to_str(&tcp_sock->src),
2936 (unsigned)ntohs(tcp_sock->src.ip6.sin6_port), client_id, client->pid));
2939 DEBUG(DEBUG_ERR,(__location__ " Unknown family %d\n", addr.sa.sa_family));
2943 /* tell all nodes about this tcp connection */
2944 ret = ctdb_daemon_send_control(ctdb, CTDB_BROADCAST_CONNECTED, 0,
2945 CTDB_CONTROL_TCP_ADD,
2946 0, CTDB_CTRL_FLAG_NOREPLY, data, NULL, NULL);
2948 DEBUG(DEBUG_ERR,(__location__ " Failed to send CTDB_CONTROL_TCP_ADD\n"));
2956 find a tcp address on a list
2958 static struct ctdb_connection *ctdb_tcp_find(struct ctdb_tcp_array *array,
2959 struct ctdb_connection *tcp)
2963 if (array == NULL) {
2967 for (i=0;i<array->num;i++) {
2968 if (ctdb_same_sockaddr(&array->connections[i].src, &tcp->src) &&
2969 ctdb_same_sockaddr(&array->connections[i].dst, &tcp->dst)) {
2970 return &array->connections[i];
2979 called by a daemon to inform us of a TCP connection that one of its
2980 clients managing that should tickled with an ACK when IP takeover is
2983 int32_t ctdb_control_tcp_add(struct ctdb_context *ctdb, TDB_DATA indata, bool tcp_update_needed)
2985 struct ctdb_connection *p = (struct ctdb_connection *)indata.dptr;
2986 struct ctdb_tcp_array *tcparray;
2987 struct ctdb_connection tcp;
2988 struct ctdb_vnn *vnn;
2990 /* If we don't have public IPs, tickles are useless */
2991 if (ctdb->vnn == NULL) {
2995 vnn = find_public_ip_vnn(ctdb, &p->dst);
2997 DEBUG(DEBUG_INFO,(__location__ " got TCP_ADD control for an address which is not a public address '%s'\n",
2998 ctdb_addr_to_str(&p->dst)));
3004 tcparray = vnn->tcp_array;
3006 /* If this is the first tickle */
3007 if (tcparray == NULL) {
3008 tcparray = talloc(vnn, struct ctdb_tcp_array);
3009 CTDB_NO_MEMORY(ctdb, tcparray);
3010 vnn->tcp_array = tcparray;
3013 tcparray->connections = talloc_size(tcparray, sizeof(struct ctdb_connection));
3014 CTDB_NO_MEMORY(ctdb, tcparray->connections);
3016 tcparray->connections[tcparray->num].src = p->src;
3017 tcparray->connections[tcparray->num].dst = p->dst;
3020 if (tcp_update_needed) {
3021 vnn->tcp_update_needed = true;
3027 /* Do we already have this tickle ?*/
3030 if (ctdb_tcp_find(tcparray, &tcp) != NULL) {
3031 DEBUG(DEBUG_DEBUG,("Already had tickle info for %s:%u for vnn:%u\n",
3032 ctdb_addr_to_str(&tcp.dst),
3033 ntohs(tcp.dst.ip.sin_port),
3038 /* A new tickle, we must add it to the array */
3039 tcparray->connections = talloc_realloc(tcparray, tcparray->connections,
3040 struct ctdb_connection,
3042 CTDB_NO_MEMORY(ctdb, tcparray->connections);
3044 tcparray->connections[tcparray->num].src = p->src;
3045 tcparray->connections[tcparray->num].dst = p->dst;
3048 DEBUG(DEBUG_INFO,("Added tickle info for %s:%u from vnn %u\n",
3049 ctdb_addr_to_str(&tcp.dst),
3050 ntohs(tcp.dst.ip.sin_port),
3053 if (tcp_update_needed) {
3054 vnn->tcp_update_needed = true;
3062 called by a daemon to inform us of a TCP connection that one of its
3063 clients managing that should tickled with an ACK when IP takeover is
3066 static void ctdb_remove_connection(struct ctdb_context *ctdb, struct ctdb_connection *conn)
3068 struct ctdb_connection *tcpp;
3069 struct ctdb_vnn *vnn = find_public_ip_vnn(ctdb, &conn->dst);
3072 DEBUG(DEBUG_ERR,(__location__ " unable to find public address %s\n",
3073 ctdb_addr_to_str(&conn->dst)));
3077 /* if the array is empty we cant remove it
3078 and we don't need to do anything
3080 if (vnn->tcp_array == NULL) {
3081 DEBUG(DEBUG_INFO,("Trying to remove tickle that doesnt exist (array is empty) %s:%u\n",
3082 ctdb_addr_to_str(&conn->dst),
3083 ntohs(conn->dst.ip.sin_port)));
3088 /* See if we know this connection
3089 if we don't know this connection then we dont need to do anything
3091 tcpp = ctdb_tcp_find(vnn->tcp_array, conn);
3093 DEBUG(DEBUG_INFO,("Trying to remove tickle that doesnt exist %s:%u\n",
3094 ctdb_addr_to_str(&conn->dst),
3095 ntohs(conn->dst.ip.sin_port)));
3100 /* We need to remove this entry from the array.
3101 Instead of allocating a new array and copying data to it
3102 we cheat and just copy the last entry in the existing array
3103 to the entry that is to be removed and just shring the
3106 *tcpp = vnn->tcp_array->connections[vnn->tcp_array->num - 1];
3107 vnn->tcp_array->num--;
3109 /* If we deleted the last entry we also need to remove the entire array
3111 if (vnn->tcp_array->num == 0) {
3112 talloc_free(vnn->tcp_array);
3113 vnn->tcp_array = NULL;
3116 vnn->tcp_update_needed = true;
3118 DEBUG(DEBUG_INFO,("Removed tickle info for %s:%u\n",
3119 ctdb_addr_to_str(&conn->src),
3120 ntohs(conn->src.ip.sin_port)));
3125 called by a daemon to inform us of a TCP connection that one of its
3126 clients used are no longer needed in the tickle database
3128 int32_t ctdb_control_tcp_remove(struct ctdb_context *ctdb, TDB_DATA indata)
3130 struct ctdb_connection *conn = (struct ctdb_connection *)indata.dptr;
3132 /* If we don't have public IPs, tickles are useless */
3133 if (ctdb->vnn == NULL) {
3137 ctdb_remove_connection(ctdb, conn);
3144 Called when another daemon starts - causes all tickles for all
3145 public addresses we are serving to be sent to the new node on the
3146 next check. This actually causes the next scheduled call to
3147 tdb_update_tcp_tickles() to update all nodes. This is simple and
3148 doesn't require careful error handling.
3150 int32_t ctdb_control_startup(struct ctdb_context *ctdb, uint32_t pnn)
3152 struct ctdb_vnn *vnn;
3154 DEBUG(DEBUG_INFO, ("Received startup control from node %lu\n",
3155 (unsigned long) pnn));
3157 for (vnn = ctdb->vnn; vnn != NULL; vnn = vnn->next) {
3158 vnn->tcp_update_needed = true;
3166 called when a client structure goes away - hook to remove
3167 elements from the tcp_list in all daemons
3169 void ctdb_takeover_client_destructor_hook(struct ctdb_client *client)
3171 while (client->tcp_list) {
3172 struct ctdb_tcp_list *tcp = client->tcp_list;
3173 DLIST_REMOVE(client->tcp_list, tcp);
3174 ctdb_remove_connection(client->ctdb, &tcp->connection);
3179 void ctdb_release_all_ips(struct ctdb_context *ctdb)
3181 struct ctdb_vnn *vnn;
3184 if (ctdb->tunable.disable_ip_failover == 1) {
3188 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
3189 if (!ctdb_sys_have_ip(&vnn->public_address)) {
3190 ctdb_vnn_unassign_iface(ctdb, vnn);
3197 /* Don't allow multiple releases at once. Some code,
3198 * particularly ctdb_tickle_sentenced_connections() is
3200 if (vnn->update_in_flight) {
3201 DEBUG(DEBUG_WARNING,
3203 " Not releasing IP %s/%u on interface %s, an update is already in progess\n",
3204 ctdb_addr_to_str(&vnn->public_address),
3205 vnn->public_netmask_bits,
3206 ctdb_vnn_iface_string(vnn)));
3209 vnn->update_in_flight = true;
3211 DEBUG(DEBUG_INFO,("Release of IP %s/%u on interface %s node:-1\n",
3212 ctdb_addr_to_str(&vnn->public_address),
3213 vnn->public_netmask_bits,
3214 ctdb_vnn_iface_string(vnn)));
3216 ctdb_event_script_args(ctdb, CTDB_EVENT_RELEASE_IP, "%s %s %u",
3217 ctdb_vnn_iface_string(vnn),
3218 ctdb_addr_to_str(&vnn->public_address),
3219 vnn->public_netmask_bits);
3220 release_kill_clients(ctdb, &vnn->public_address);
3221 ctdb_vnn_unassign_iface(ctdb, vnn);
3222 vnn->update_in_flight = false;
3226 DEBUG(DEBUG_NOTICE,(__location__ " Released %d public IPs\n", count));
3231 get list of public IPs
3233 int32_t ctdb_control_get_public_ips(struct ctdb_context *ctdb,
3234 struct ctdb_req_control_old *c, TDB_DATA *outdata)
3237 struct ctdb_public_ip_list_old *ips;
3238 struct ctdb_vnn *vnn;
3239 bool only_available = false;
3241 if (c->flags & CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE) {
3242 only_available = true;
3245 /* count how many public ip structures we have */
3247 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
3251 len = offsetof(struct ctdb_public_ip_list_old, ips) +
3252 num*sizeof(struct ctdb_public_ip);
3253 ips = talloc_zero_size(outdata, len);
3254 CTDB_NO_MEMORY(ctdb, ips);
3257 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
3258 if (only_available && !ctdb_vnn_available(ctdb, vnn)) {
3261 ips->ips[i].pnn = vnn->pnn;
3262 ips->ips[i].addr = vnn->public_address;
3266 len = offsetof(struct ctdb_public_ip_list_old, ips) +
3267 i*sizeof(struct ctdb_public_ip);
3269 outdata->dsize = len;
3270 outdata->dptr = (uint8_t *)ips;
3276 int32_t ctdb_control_get_public_ip_info(struct ctdb_context *ctdb,
3277 struct ctdb_req_control_old *c,
3282 ctdb_sock_addr *addr;
3283 struct ctdb_public_ip_info_old *info;
3284 struct ctdb_vnn *vnn;
3286 addr = (ctdb_sock_addr *)indata.dptr;
3288 vnn = find_public_ip_vnn(ctdb, addr);
3290 /* if it is not a public ip it could be our 'single ip' */
3291 if (ctdb->single_ip_vnn) {
3292 if (ctdb_same_ip(&ctdb->single_ip_vnn->public_address, addr)) {
3293 vnn = ctdb->single_ip_vnn;
3298 DEBUG(DEBUG_ERR,(__location__ " Could not get public ip info, "
3299 "'%s'not a public address\n",
3300 ctdb_addr_to_str(addr)));
3304 /* count how many public ip structures we have */
3306 for (;vnn->ifaces[num];) {
3310 len = offsetof(struct ctdb_public_ip_info_old, ifaces) +
3311 num*sizeof(struct ctdb_iface);
3312 info = talloc_zero_size(outdata, len);
3313 CTDB_NO_MEMORY(ctdb, info);
3315 info->ip.addr = vnn->public_address;
3316 info->ip.pnn = vnn->pnn;
3317 info->active_idx = 0xFFFFFFFF;
3319 for (i=0; vnn->ifaces[i]; i++) {
3320 struct ctdb_interface *cur;
3322 cur = ctdb_find_iface(ctdb, vnn->ifaces[i]);
3324 DEBUG(DEBUG_CRIT, (__location__ " internal error iface[%s] unknown\n",
3328 if (vnn->iface == cur) {
3329 info->active_idx = i;
3331 strncpy(info->ifaces[i].name, cur->name, sizeof(info->ifaces[i].name)-1);
3332 info->ifaces[i].link_state = cur->link_up;
3333 info->ifaces[i].references = cur->references;
3336 len = offsetof(struct ctdb_public_ip_info_old, ifaces) +
3337 i*sizeof(struct ctdb_iface);
3339 outdata->dsize = len;
3340 outdata->dptr = (uint8_t *)info;
3345 int32_t ctdb_control_get_ifaces(struct ctdb_context *ctdb,
3346 struct ctdb_req_control_old *c,
3350 struct ctdb_iface_list_old *ifaces;
3351 struct ctdb_interface *cur;
3353 /* count how many public ip structures we have */
3355 for (cur=ctdb->ifaces;cur;cur=cur->next) {
3359 len = offsetof(struct ctdb_iface_list_old, ifaces) +
3360 num*sizeof(struct ctdb_iface);
3361 ifaces = talloc_zero_size(outdata, len);
3362 CTDB_NO_MEMORY(ctdb, ifaces);
3365 for (cur=ctdb->ifaces;cur;cur=cur->next) {
3366 strcpy(ifaces->ifaces[i].name, cur->name);
3367 ifaces->ifaces[i].link_state = cur->link_up;
3368 ifaces->ifaces[i].references = cur->references;
3372 len = offsetof(struct ctdb_iface_list_old, ifaces) +
3373 i*sizeof(struct ctdb_iface);
3375 outdata->dsize = len;
3376 outdata->dptr = (uint8_t *)ifaces;
3381 int32_t ctdb_control_set_iface_link(struct ctdb_context *ctdb,
3382 struct ctdb_req_control_old *c,
3385 struct ctdb_iface *info;
3386 struct ctdb_interface *iface;
3387 bool link_up = false;
3389 info = (struct ctdb_iface *)indata.dptr;
3391 if (info->name[CTDB_IFACE_SIZE] != '\0') {
3392 int len = strnlen(info->name, CTDB_IFACE_SIZE);
3393 DEBUG(DEBUG_ERR, (__location__ " name[%*.*s] not terminated\n",
3394 len, len, info->name));
3398 switch (info->link_state) {
3406 DEBUG(DEBUG_ERR, (__location__ " link_state[%u] invalid\n",
3407 (unsigned int)info->link_state));
3411 if (info->references != 0) {
3412 DEBUG(DEBUG_ERR, (__location__ " references[%u] should be 0\n",
3413 (unsigned int)info->references));
3417 iface = ctdb_find_iface(ctdb, info->name);
3418 if (iface == NULL) {
3422 if (link_up == iface->link_up) {
3426 DEBUG(iface->link_up?DEBUG_ERR:DEBUG_NOTICE,
3427 ("iface[%s] has changed it's link status %s => %s\n",
3429 iface->link_up?"up":"down",
3430 link_up?"up":"down"));
3432 iface->link_up = link_up;
3438 structure containing the listening socket and the list of tcp connections
3439 that the ctdb daemon is to kill
3441 struct ctdb_kill_tcp {
3442 struct ctdb_vnn *vnn;
3443 struct ctdb_context *ctdb;
3445 struct tevent_fd *fde;
3446 trbt_tree_t *connections;
3451 a tcp connection that is to be killed
3453 struct ctdb_killtcp_con {
3454 ctdb_sock_addr src_addr;
3455 ctdb_sock_addr dst_addr;
3457 struct ctdb_kill_tcp *killtcp;
3460 /* this function is used to create a key to represent this socketpair
3461 in the killtcp tree.
3462 this key is used to insert and lookup matching socketpairs that are
3463 to be tickled and RST
3465 #define KILLTCP_KEYLEN 10
3466 static uint32_t *killtcp_key(ctdb_sock_addr *src, ctdb_sock_addr *dst)
3468 static uint32_t key[KILLTCP_KEYLEN];
3470 bzero(key, sizeof(key));
3472 if (src->sa.sa_family != dst->sa.sa_family) {
3473 DEBUG(DEBUG_ERR, (__location__ " ERROR, different families passed :%u vs %u\n", src->sa.sa_family, dst->sa.sa_family));
3477 switch (src->sa.sa_family) {
3479 key[0] = dst->ip.sin_addr.s_addr;
3480 key[1] = src->ip.sin_addr.s_addr;
3481 key[2] = dst->ip.sin_port;
3482 key[3] = src->ip.sin_port;
3485 uint32_t *dst6_addr32 =
3486 (uint32_t *)&(dst->ip6.sin6_addr.s6_addr);
3487 uint32_t *src6_addr32 =
3488 (uint32_t *)&(src->ip6.sin6_addr.s6_addr);
3489 key[0] = dst6_addr32[3];
3490 key[1] = src6_addr32[3];
3491 key[2] = dst6_addr32[2];
3492 key[3] = src6_addr32[2];
3493 key[4] = dst6_addr32[1];
3494 key[5] = src6_addr32[1];
3495 key[6] = dst6_addr32[0];
3496 key[7] = src6_addr32[0];
3497 key[8] = dst->ip6.sin6_port;
3498 key[9] = src->ip6.sin6_port;
3502 DEBUG(DEBUG_ERR, (__location__ " ERROR, unknown family passed :%u\n", src->sa.sa_family));
3510 called when we get a read event on the raw socket
3512 static void capture_tcp_handler(struct tevent_context *ev,
3513 struct tevent_fd *fde,
3514 uint16_t flags, void *private_data)
3516 struct ctdb_kill_tcp *killtcp = talloc_get_type(private_data, struct ctdb_kill_tcp);
3517 struct ctdb_killtcp_con *con;
3518 ctdb_sock_addr src, dst;
3519 uint32_t ack_seq, seq;
3521 if (!(flags & TEVENT_FD_READ)) {
3525 if (ctdb_sys_read_tcp_packet(killtcp->capture_fd,
3526 killtcp->private_data,
3528 &ack_seq, &seq) != 0) {
3529 /* probably a non-tcp ACK packet */
3533 /* check if we have this guy in our list of connections
3536 con = trbt_lookuparray32(killtcp->connections,
3537 KILLTCP_KEYLEN, killtcp_key(&src, &dst));
3539 /* no this was some other packet we can just ignore */
3543 /* This one has been tickled !
3544 now reset him and remove him from the list.
3546 DEBUG(DEBUG_INFO, ("sending a tcp reset to kill connection :%d -> %s:%d\n",
3547 ntohs(con->dst_addr.ip.sin_port),
3548 ctdb_addr_to_str(&con->src_addr),
3549 ntohs(con->src_addr.ip.sin_port)));
3551 ctdb_sys_send_tcp(&con->dst_addr, &con->src_addr, ack_seq, seq, 1);
3556 /* when traversing the list of all tcp connections to send tickle acks to
3557 (so that we can capture the ack coming back and kill the connection
3559 this callback is called for each connection we are currently trying to kill
3561 static int tickle_connection_traverse(void *param, void *data)
3563 struct ctdb_killtcp_con *con = talloc_get_type(data, struct ctdb_killtcp_con);
3565 /* have tried too many times, just give up */
3566 if (con->count >= 5) {
3567 /* can't delete in traverse: reparent to delete_cons */
3568 talloc_steal(param, con);
3572 /* othervise, try tickling it again */
3575 (ctdb_sock_addr *)&con->dst_addr,
3576 (ctdb_sock_addr *)&con->src_addr,
3583 called every second until all sentenced connections have been reset
3585 static void ctdb_tickle_sentenced_connections(struct tevent_context *ev,
3586 struct tevent_timer *te,
3587 struct timeval t, void *private_data)
3589 struct ctdb_kill_tcp *killtcp = talloc_get_type(private_data, struct ctdb_kill_tcp);
3590 void *delete_cons = talloc_new(NULL);
3592 /* loop over all connections sending tickle ACKs */
3593 trbt_traversearray32(killtcp->connections, KILLTCP_KEYLEN, tickle_connection_traverse, delete_cons);
3595 /* now we've finished traverse, it's safe to do deletion. */
3596 talloc_free(delete_cons);
3598 /* If there are no more connections to kill we can remove the
3599 entire killtcp structure
3601 if ( (killtcp->connections == NULL) ||
3602 (killtcp->connections->root == NULL) ) {
3603 talloc_free(killtcp);
3607 /* try tickling them again in a seconds time
3609 tevent_add_timer(killtcp->ctdb->ev, killtcp,
3610 timeval_current_ofs(1, 0),
3611 ctdb_tickle_sentenced_connections, killtcp);
3615 destroy the killtcp structure
3617 static int ctdb_killtcp_destructor(struct ctdb_kill_tcp *killtcp)
3619 struct ctdb_vnn *tmpvnn;
3621 /* verify that this vnn is still active */
3622 for (tmpvnn = killtcp->ctdb->vnn; tmpvnn; tmpvnn = tmpvnn->next) {
3623 if (tmpvnn == killtcp->vnn) {
3628 if (tmpvnn == NULL) {
3632 if (killtcp->vnn->killtcp != killtcp) {
3636 killtcp->vnn->killtcp = NULL;
3642 /* nothing fancy here, just unconditionally replace any existing
3643 connection structure with the new one.
3645 don't even free the old one if it did exist, that one is talloc_stolen
3646 by the same node in the tree anyway and will be deleted when the new data
3649 static void *add_killtcp_callback(void *parm, void *data)
3655 add a tcp socket to the list of connections we want to RST
3657 static int ctdb_killtcp_add_connection(struct ctdb_context *ctdb,
3661 ctdb_sock_addr src, dst;
3662 struct ctdb_kill_tcp *killtcp;
3663 struct ctdb_killtcp_con *con;
3664 struct ctdb_vnn *vnn;
3666 ctdb_canonicalize_ip(s, &src);
3667 ctdb_canonicalize_ip(d, &dst);
3669 vnn = find_public_ip_vnn(ctdb, &dst);
3671 vnn = find_public_ip_vnn(ctdb, &src);
3674 /* if it is not a public ip it could be our 'single ip' */
3675 if (ctdb->single_ip_vnn) {
3676 if (ctdb_same_ip(&ctdb->single_ip_vnn->public_address, &dst)) {
3677 vnn = ctdb->single_ip_vnn;
3682 DEBUG(DEBUG_ERR,(__location__ " Could not killtcp, not a public address\n"));
3686 killtcp = vnn->killtcp;
3688 /* If this is the first connection to kill we must allocate
3691 if (killtcp == NULL) {
3692 killtcp = talloc_zero(vnn, struct ctdb_kill_tcp);
3693 CTDB_NO_MEMORY(ctdb, killtcp);
3696 killtcp->ctdb = ctdb;
3697 killtcp->capture_fd = -1;
3698 killtcp->connections = trbt_create(killtcp, 0);
3700 vnn->killtcp = killtcp;
3701 talloc_set_destructor(killtcp, ctdb_killtcp_destructor);
3706 /* create a structure that describes this connection we want to
3707 RST and store it in killtcp->connections
3709 con = talloc(killtcp, struct ctdb_killtcp_con);
3710 CTDB_NO_MEMORY(ctdb, con);
3711 con->src_addr = src;
3712 con->dst_addr = dst;
3714 con->killtcp = killtcp;
3717 trbt_insertarray32_callback(killtcp->connections,
3718 KILLTCP_KEYLEN, killtcp_key(&con->dst_addr, &con->src_addr),
3719 add_killtcp_callback, con);
3722 If we don't have a socket to listen on yet we must create it
3724 if (killtcp->capture_fd == -1) {
3725 const char *iface = ctdb_vnn_iface_string(vnn);
3726 killtcp->capture_fd = ctdb_sys_open_capture_socket(iface, &killtcp->private_data);
3727 if (killtcp->capture_fd == -1) {
3728 DEBUG(DEBUG_CRIT,(__location__ " Failed to open capturing "
3729 "socket on iface '%s' for killtcp (%s)\n",
3730 iface, strerror(errno)));
3736 if (killtcp->fde == NULL) {
3737 killtcp->fde = tevent_add_fd(ctdb->ev, killtcp,
3738 killtcp->capture_fd,
3740 capture_tcp_handler, killtcp);
3741 tevent_fd_set_auto_close(killtcp->fde);
3743 /* We also need to set up some events to tickle all these connections
3744 until they are all reset
3746 tevent_add_timer(ctdb->ev, killtcp, timeval_current_ofs(1, 0),
3747 ctdb_tickle_sentenced_connections, killtcp);
3750 /* tickle him once now */
3759 talloc_free(vnn->killtcp);
3760 vnn->killtcp = NULL;
3765 kill a TCP connection.
3767 int32_t ctdb_control_kill_tcp(struct ctdb_context *ctdb, TDB_DATA indata)
3769 struct ctdb_connection *killtcp = (struct ctdb_connection *)indata.dptr;
3771 return ctdb_killtcp_add_connection(ctdb, &killtcp->src, &killtcp->dst);
3775 called by a daemon to inform us of the entire list of TCP tickles for
3776 a particular public address.
3777 this control should only be sent by the node that is currently serving
3778 that public address.
3780 int32_t ctdb_control_set_tcp_tickle_list(struct ctdb_context *ctdb, TDB_DATA indata)
3782 struct ctdb_tickle_list_old *list = (struct ctdb_tickle_list_old *)indata.dptr;
3783 struct ctdb_tcp_array *tcparray;
3784 struct ctdb_vnn *vnn;
3786 /* We must at least have tickles.num or else we cant verify the size
3787 of the received data blob
3789 if (indata.dsize < offsetof(struct ctdb_tickle_list_old, connections)) {
3790 DEBUG(DEBUG_ERR,("Bad indata in ctdb_tickle_list. Not enough data for the tickle.num field\n"));
3794 /* verify that the size of data matches what we expect */
3795 if (indata.dsize < offsetof(struct ctdb_tickle_list_old, connections)
3796 + sizeof(struct ctdb_connection) * list->num) {
3797 DEBUG(DEBUG_ERR,("Bad indata in ctdb_tickle_list\n"));
3801 DEBUG(DEBUG_INFO, ("Received tickle update for public address %s\n",
3802 ctdb_addr_to_str(&list->addr)));
3804 vnn = find_public_ip_vnn(ctdb, &list->addr);
3806 DEBUG(DEBUG_INFO,(__location__ " Could not set tcp tickle list, '%s' is not a public address\n",
3807 ctdb_addr_to_str(&list->addr)));
3812 /* remove any old ticklelist we might have */
3813 talloc_free(vnn->tcp_array);
3814 vnn->tcp_array = NULL;
3816 tcparray = talloc(vnn, struct ctdb_tcp_array);
3817 CTDB_NO_MEMORY(ctdb, tcparray);
3819 tcparray->num = list->num;
3821 tcparray->connections = talloc_array(tcparray, struct ctdb_connection, tcparray->num);
3822 CTDB_NO_MEMORY(ctdb, tcparray->connections);
3824 memcpy(tcparray->connections, &list->connections[0],
3825 sizeof(struct ctdb_connection)*tcparray->num);
3827 /* We now have a new fresh tickle list array for this vnn */
3828 vnn->tcp_array = tcparray;
3834 called to return the full list of tickles for the puclic address associated
3835 with the provided vnn
3837 int32_t ctdb_control_get_tcp_tickle_list(struct ctdb_context *ctdb, TDB_DATA indata, TDB_DATA *outdata)
3839 ctdb_sock_addr *addr = (ctdb_sock_addr *)indata.dptr;
3840 struct ctdb_tickle_list_old *list;
3841 struct ctdb_tcp_array *tcparray;
3843 struct ctdb_vnn *vnn;
3845 vnn = find_public_ip_vnn(ctdb, addr);
3847 DEBUG(DEBUG_ERR,(__location__ " Could not get tcp tickle list, '%s' is not a public address\n",
3848 ctdb_addr_to_str(addr)));
3853 tcparray = vnn->tcp_array;
3855 num = tcparray->num;
3860 outdata->dsize = offsetof(struct ctdb_tickle_list_old, connections)
3861 + sizeof(struct ctdb_connection) * num;
3863 outdata->dptr = talloc_size(outdata, outdata->dsize);
3864 CTDB_NO_MEMORY(ctdb, outdata->dptr);
3865 list = (struct ctdb_tickle_list_old *)outdata->dptr;
3870 memcpy(&list->connections[0], tcparray->connections,
3871 sizeof(struct ctdb_connection) * num);
3879 set the list of all tcp tickles for a public address
3881 static int ctdb_send_set_tcp_tickles_for_ip(struct ctdb_context *ctdb,
3882 ctdb_sock_addr *addr,
3883 struct ctdb_tcp_array *tcparray)
3887 struct ctdb_tickle_list_old *list;
3890 num = tcparray->num;
3895 data.dsize = offsetof(struct ctdb_tickle_list_old, connections) +
3896 sizeof(struct ctdb_connection) * num;
3897 data.dptr = talloc_size(ctdb, data.dsize);
3898 CTDB_NO_MEMORY(ctdb, data.dptr);
3900 list = (struct ctdb_tickle_list_old *)data.dptr;
3904 memcpy(&list->connections[0], tcparray->connections, sizeof(struct ctdb_connection) * num);
3907 ret = ctdb_daemon_send_control(ctdb, CTDB_BROADCAST_ALL, 0,
3908 CTDB_CONTROL_SET_TCP_TICKLE_LIST,
3909 0, CTDB_CTRL_FLAG_NOREPLY, data, NULL, NULL);
3911 DEBUG(DEBUG_ERR,(__location__ " ctdb_control for set tcp tickles failed\n"));
3915 talloc_free(data.dptr);
3922 perform tickle updates if required
3924 static void ctdb_update_tcp_tickles(struct tevent_context *ev,
3925 struct tevent_timer *te,
3926 struct timeval t, void *private_data)
3928 struct ctdb_context *ctdb = talloc_get_type(private_data, struct ctdb_context);
3930 struct ctdb_vnn *vnn;
3932 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
3933 /* we only send out updates for public addresses that
3936 if (ctdb->pnn != vnn->pnn) {
3939 /* We only send out the updates if we need to */
3940 if (!vnn->tcp_update_needed) {
3943 ret = ctdb_send_set_tcp_tickles_for_ip(ctdb,
3944 &vnn->public_address,
3947 DEBUG(DEBUG_ERR,("Failed to send the tickle update for public address %s\n",
3948 ctdb_addr_to_str(&vnn->public_address)));
3951 ("Sent tickle update for public address %s\n",
3952 ctdb_addr_to_str(&vnn->public_address)));
3953 vnn->tcp_update_needed = false;
3957 tevent_add_timer(ctdb->ev, ctdb->tickle_update_context,
3958 timeval_current_ofs(ctdb->tunable.tickle_update_interval, 0),
3959 ctdb_update_tcp_tickles, ctdb);
3963 start periodic update of tcp tickles
3965 void ctdb_start_tcp_tickle_update(struct ctdb_context *ctdb)
3967 ctdb->tickle_update_context = talloc_new(ctdb);
3969 tevent_add_timer(ctdb->ev, ctdb->tickle_update_context,
3970 timeval_current_ofs(ctdb->tunable.tickle_update_interval, 0),
3971 ctdb_update_tcp_tickles, ctdb);
3977 struct control_gratious_arp {
3978 struct ctdb_context *ctdb;
3979 ctdb_sock_addr addr;
3985 send a control_gratuitous arp
3987 static void send_gratious_arp(struct tevent_context *ev,
3988 struct tevent_timer *te,
3989 struct timeval t, void *private_data)
3992 struct control_gratious_arp *arp = talloc_get_type(private_data,
3993 struct control_gratious_arp);
3995 ret = ctdb_sys_send_arp(&arp->addr, arp->iface);
3997 DEBUG(DEBUG_ERR,(__location__ " sending of gratious arp on iface '%s' failed (%s)\n",
3998 arp->iface, strerror(errno)));
4003 if (arp->count == CTDB_ARP_REPEAT) {
4008 tevent_add_timer(arp->ctdb->ev, arp,
4009 timeval_current_ofs(CTDB_ARP_INTERVAL, 0),
4010 send_gratious_arp, arp);
4017 int32_t ctdb_control_send_gratious_arp(struct ctdb_context *ctdb, TDB_DATA indata)
4019 struct ctdb_addr_info_old *gratious_arp = (struct ctdb_addr_info_old *)indata.dptr;
4020 struct control_gratious_arp *arp;
4022 /* verify the size of indata */
4023 if (indata.dsize < offsetof(struct ctdb_addr_info_old, iface)) {
4024 DEBUG(DEBUG_ERR,(__location__ " Too small indata to hold a ctdb_control_gratious_arp structure. Got %u require %u bytes\n",
4025 (unsigned)indata.dsize,
4026 (unsigned)offsetof(struct ctdb_addr_info_old, iface)));
4030 ( offsetof(struct ctdb_addr_info_old, iface)
4031 + gratious_arp->len ) ){
4033 DEBUG(DEBUG_ERR,(__location__ " Wrong size of indata. Was %u bytes "
4034 "but should be %u bytes\n",
4035 (unsigned)indata.dsize,
4036 (unsigned)(offsetof(struct ctdb_addr_info_old, iface)+gratious_arp->len)));
4041 arp = talloc(ctdb, struct control_gratious_arp);
4042 CTDB_NO_MEMORY(ctdb, arp);
4045 arp->addr = gratious_arp->addr;
4046 arp->iface = talloc_strdup(arp, gratious_arp->iface);
4047 CTDB_NO_MEMORY(ctdb, arp->iface);
4050 tevent_add_timer(arp->ctdb->ev, arp,
4051 timeval_zero(), send_gratious_arp, arp);
4056 int32_t ctdb_control_add_public_address(struct ctdb_context *ctdb, TDB_DATA indata)
4058 struct ctdb_addr_info_old *pub = (struct ctdb_addr_info_old *)indata.dptr;
4061 /* verify the size of indata */
4062 if (indata.dsize < offsetof(struct ctdb_addr_info_old, iface)) {
4063 DEBUG(DEBUG_ERR,(__location__ " Too small indata to hold a ctdb_addr_info structure\n"));
4067 ( offsetof(struct ctdb_addr_info_old, iface)
4070 DEBUG(DEBUG_ERR,(__location__ " Wrong size of indata. Was %u bytes "
4071 "but should be %u bytes\n",
4072 (unsigned)indata.dsize,
4073 (unsigned)(offsetof(struct ctdb_addr_info_old, iface)+pub->len)));
4077 DEBUG(DEBUG_NOTICE,("Add IP %s\n", ctdb_addr_to_str(&pub->addr)));
4079 ret = ctdb_add_public_address(ctdb, &pub->addr, pub->mask, &pub->iface[0], true);
4082 DEBUG(DEBUG_ERR,(__location__ " Failed to add public address\n"));
4089 struct delete_ip_callback_state {
4090 struct ctdb_req_control_old *c;
4094 called when releaseip event finishes for del_public_address
4096 static void delete_ip_callback(struct ctdb_context *ctdb,
4097 int32_t status, TDB_DATA data,
4098 const char *errormsg,
4101 struct delete_ip_callback_state *state =
4102 talloc_get_type(private_data, struct delete_ip_callback_state);
4104 /* If release failed then fail. */
4105 ctdb_request_control_reply(ctdb, state->c, NULL, status, errormsg);
4106 talloc_free(private_data);
4109 int32_t ctdb_control_del_public_address(struct ctdb_context *ctdb,
4110 struct ctdb_req_control_old *c,
4111 TDB_DATA indata, bool *async_reply)
4113 struct ctdb_addr_info_old *pub = (struct ctdb_addr_info_old *)indata.dptr;
4114 struct ctdb_vnn *vnn;
4116 /* verify the size of indata */
4117 if (indata.dsize < offsetof(struct ctdb_addr_info_old, iface)) {
4118 DEBUG(DEBUG_ERR,(__location__ " Too small indata to hold a ctdb_addr_info structure\n"));
4122 ( offsetof(struct ctdb_addr_info_old, iface)
4125 DEBUG(DEBUG_ERR,(__location__ " Wrong size of indata. Was %u bytes "
4126 "but should be %u bytes\n",
4127 (unsigned)indata.dsize,
4128 (unsigned)(offsetof(struct ctdb_addr_info_old, iface)+pub->len)));
4132 DEBUG(DEBUG_NOTICE,("Delete IP %s\n", ctdb_addr_to_str(&pub->addr)));
4134 /* walk over all public addresses until we find a match */
4135 for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
4136 if (ctdb_same_ip(&vnn->public_address, &pub->addr)) {
4137 if (vnn->pnn == ctdb->pnn) {
4138 struct delete_ip_callback_state *state;
4139 struct ctdb_public_ip *ip;
4143 vnn->delete_pending = true;
4145 state = talloc(ctdb,
4146 struct delete_ip_callback_state);
4147 CTDB_NO_MEMORY(ctdb, state);
4150 ip = talloc(state, struct ctdb_public_ip);
4153 (__location__ " Out of memory\n"));
4158 ip->addr = pub->addr;
4160 data.dsize = sizeof(struct ctdb_public_ip);
4161 data.dptr = (unsigned char *)ip;
4163 ret = ctdb_daemon_send_control(ctdb,
4166 CTDB_CONTROL_RELEASE_IP,
4173 (__location__ "Unable to send "
4174 "CTDB_CONTROL_RELEASE_IP\n"));
4179 state->c = talloc_steal(state, c);
4180 *async_reply = true;
4182 /* This IP is not hosted on the
4183 * current node so just delete it
4185 do_delete_ip(ctdb, vnn);
4192 DEBUG(DEBUG_ERR,("Delete IP of unknown public IP address %s\n",
4193 ctdb_addr_to_str(&pub->addr)));
4198 struct ipreallocated_callback_state {
4199 struct ctdb_req_control_old *c;
4202 static void ctdb_ipreallocated_callback(struct ctdb_context *ctdb,
4203 int status, void *p)
4205 struct ipreallocated_callback_state *state =
4206 talloc_get_type(p, struct ipreallocated_callback_state);
4210 (" \"ipreallocated\" event script failed (status %d)\n",
4212 if (status == -ETIME) {
4213 ctdb_ban_self(ctdb);
4217 ctdb_request_control_reply(ctdb, state->c, NULL, status, NULL);
4221 /* A control to run the ipreallocated event */
4222 int32_t ctdb_control_ipreallocated(struct ctdb_context *ctdb,
4223 struct ctdb_req_control_old *c,
4227 struct ipreallocated_callback_state *state;
4229 state = talloc(ctdb, struct ipreallocated_callback_state);
4230 CTDB_NO_MEMORY(ctdb, state);
4232 DEBUG(DEBUG_INFO,(__location__ " Running \"ipreallocated\" event\n"));
4234 ret = ctdb_event_script_callback(ctdb, state,
4235 ctdb_ipreallocated_callback, state,
4236 CTDB_EVENT_IPREALLOCATED,
4240 DEBUG(DEBUG_ERR,("Failed to run \"ipreallocated\" event \n"));
4245 /* tell the control that we will be reply asynchronously */
4246 state->c = talloc_steal(state, c);
4247 *async_reply = true;
4253 /* This function is called from the recovery daemon to verify that a remote
4254 node has the expected ip allocation.
4255 This is verified against ctdb->ip_tree
4257 static int verify_remote_ip_allocation(struct ctdb_context *ctdb,
4258 struct ctdb_public_ip_list_old *ips,
4261 struct public_ip_list *tmp_ip;
4264 if (ctdb->ip_tree == NULL) {
4265 /* don't know the expected allocation yet, assume remote node
4274 for (i=0; i<ips->num; i++) {
4275 tmp_ip = trbt_lookuparray32(ctdb->ip_tree, IP_KEYLEN, ip_key(&ips->ips[i].addr));
4276 if (tmp_ip == NULL) {
4277 DEBUG(DEBUG_ERR,("Node %u has new or unknown public IP %s\n", pnn, ctdb_addr_to_str(&ips->ips[i].addr)));
4281 if (tmp_ip->pnn == -1 || ips->ips[i].pnn == -1) {
4285 if (tmp_ip->pnn != ips->ips[i].pnn) {
4287 ("Inconsistent IP allocation - node %u thinks %s is held by node %u while it is assigned to node %u\n",
4289 ctdb_addr_to_str(&ips->ips[i].addr),
4290 ips->ips[i].pnn, tmp_ip->pnn));
4298 int update_ip_assignment_tree(struct ctdb_context *ctdb, struct ctdb_public_ip *ip)
4300 struct public_ip_list *tmp_ip;
4302 /* IP tree is never built if DisableIPFailover is set */
4303 if (ctdb->tunable.disable_ip_failover != 0) {
4307 if (ctdb->ip_tree == NULL) {
4308 DEBUG(DEBUG_ERR,("No ctdb->ip_tree yet. Failed to update ip assignment\n"));
4312 tmp_ip = trbt_lookuparray32(ctdb->ip_tree, IP_KEYLEN, ip_key(&ip->addr));
4313 if (tmp_ip == NULL) {
4314 DEBUG(DEBUG_ERR,(__location__ " Could not find record for address %s, update ip\n", ctdb_addr_to_str(&ip->addr)));
4318 DEBUG(DEBUG_NOTICE,("Updated ip assignment tree for ip : %s from node %u to node %u\n", ctdb_addr_to_str(&ip->addr), tmp_ip->pnn, ip->pnn));
4319 tmp_ip->pnn = ip->pnn;
4324 void clear_ip_assignment_tree(struct ctdb_context *ctdb)
4326 TALLOC_FREE(ctdb->ip_tree);
4329 struct ctdb_reloadips_handle {
4330 struct ctdb_context *ctdb;
4331 struct ctdb_req_control_old *c;
4335 struct tevent_fd *fde;
4338 static int ctdb_reloadips_destructor(struct ctdb_reloadips_handle *h)
4340 if (h == h->ctdb->reload_ips) {
4341 h->ctdb->reload_ips = NULL;
4344 ctdb_request_control_reply(h->ctdb, h->c, NULL, h->status, NULL);
4347 ctdb_kill(h->ctdb, h->child, SIGKILL);
4351 static void ctdb_reloadips_timeout_event(struct tevent_context *ev,
4352 struct tevent_timer *te,
4353 struct timeval t, void *private_data)
4355 struct ctdb_reloadips_handle *h = talloc_get_type(private_data, struct ctdb_reloadips_handle);
4360 static void ctdb_reloadips_child_handler(struct tevent_context *ev,
4361 struct tevent_fd *fde,
4362 uint16_t flags, void *private_data)
4364 struct ctdb_reloadips_handle *h = talloc_get_type(private_data, struct ctdb_reloadips_handle);
4369 ret = sys_read(h->fd[0], &res, 1);
4370 if (ret < 1 || res != 0) {
4371 DEBUG(DEBUG_ERR, (__location__ " Reloadips child process returned error\n"));
4379 static int ctdb_reloadips_child(struct ctdb_context *ctdb)
4381 TALLOC_CTX *mem_ctx = talloc_new(NULL);
4382 struct ctdb_public_ip_list_old *ips;
4383 struct ctdb_vnn *vnn;
4384 struct client_async_data *async_data;
4385 struct timeval timeout;
4387 struct ctdb_client_control_state *state;
4391 CTDB_NO_MEMORY(ctdb, mem_ctx);
4393 /* Read IPs from local node */
4394 ret = ctdb_ctrl_get_public_ips(ctdb, TAKEOVER_TIMEOUT(),
4395 CTDB_CURRENT_NODE, mem_ctx, &ips);
4398 ("Unable to fetch public IPs from local node\n"));
4399 talloc_free(mem_ctx);
4403 /* Read IPs file - this is safe since this is a child process */
4405 if (ctdb_set_public_addresses(ctdb, false) != 0) {
4406 DEBUG(DEBUG_ERR,("Failed to re-read public addresses file\n"));
4407 talloc_free(mem_ctx);
4411 async_data = talloc_zero(mem_ctx, struct client_async_data);
4412 CTDB_NO_MEMORY(ctdb, async_data);
4414 /* Compare IPs between node and file for IPs to be deleted */
4415 for (i = 0; i < ips->num; i++) {
4417 for (vnn = ctdb->vnn; vnn; vnn = vnn->next) {
4418 if (ctdb_same_ip(&vnn->public_address,
4419 &ips->ips[i].addr)) {
4420 /* IP is still in file */
4426 /* Delete IP ips->ips[i] */
4427 struct ctdb_addr_info_old *pub;
4430 ("IP %s no longer configured, deleting it\n",
4431 ctdb_addr_to_str(&ips->ips[i].addr)));
4433 pub = talloc_zero(mem_ctx, struct ctdb_addr_info_old);
4434 CTDB_NO_MEMORY(ctdb, pub);
4436 pub->addr = ips->ips[i].addr;
4440 timeout = TAKEOVER_TIMEOUT();
4442 data.dsize = offsetof(struct ctdb_addr_info_old,
4444 data.dptr = (uint8_t *)pub;
4446 state = ctdb_control_send(ctdb, CTDB_CURRENT_NODE, 0,
4447 CTDB_CONTROL_DEL_PUBLIC_IP,
4448 0, data, async_data,
4450 if (state == NULL) {
4453 " failed sending CTDB_CONTROL_DEL_PUBLIC_IP\n"));
4457 ctdb_client_async_add(async_data, state);
4461 /* Compare IPs between node and file for IPs to be added */
4463 for (vnn = ctdb->vnn; vnn; vnn = vnn->next) {
4464 for (i = 0; i < ips->num; i++) {
4465 if (ctdb_same_ip(&vnn->public_address,
4466 &ips->ips[i].addr)) {
4467 /* IP already on node */
4471 if (i == ips->num) {
4472 /* Add IP ips->ips[i] */
4473 struct ctdb_addr_info_old *pub;
4474 const char *ifaces = NULL;
4479 ("New IP %s configured, adding it\n",
4480 ctdb_addr_to_str(&vnn->public_address)));
4482 uint32_t pnn = ctdb_get_pnn(ctdb);
4484 data.dsize = sizeof(pnn);
4485 data.dptr = (uint8_t *)&pnn;
4487 ret = ctdb_client_send_message(
4489 CTDB_BROADCAST_CONNECTED,
4490 CTDB_SRVID_REBALANCE_NODE,
4493 DEBUG(DEBUG_WARNING,
4494 ("Failed to send message to force node reallocation - IPs may be unbalanced\n"));
4500 ifaces = vnn->ifaces[0];
4502 while (vnn->ifaces[iface] != NULL) {
4503 ifaces = talloc_asprintf(vnn, "%s,%s", ifaces,
4504 vnn->ifaces[iface]);
4508 len = strlen(ifaces) + 1;
4509 pub = talloc_zero_size(mem_ctx,
4510 offsetof(struct ctdb_addr_info_old, iface) + len);
4511 CTDB_NO_MEMORY(ctdb, pub);
4513 pub->addr = vnn->public_address;
4514 pub->mask = vnn->public_netmask_bits;
4516 memcpy(&pub->iface[0], ifaces, pub->len);
4518 timeout = TAKEOVER_TIMEOUT();
4520 data.dsize = offsetof(struct ctdb_addr_info_old,
4522 data.dptr = (uint8_t *)pub;
4524 state = ctdb_control_send(ctdb, CTDB_CURRENT_NODE, 0,
4525 CTDB_CONTROL_ADD_PUBLIC_IP,
4526 0, data, async_data,
4528 if (state == NULL) {
4531 " failed sending CTDB_CONTROL_ADD_PUBLIC_IP\n"));
4535 ctdb_client_async_add(async_data, state);
4539 if (ctdb_client_async_wait(ctdb, async_data) != 0) {
4540 DEBUG(DEBUG_ERR,(__location__ " Add/delete IPs failed\n"));
4544 talloc_free(mem_ctx);
4548 talloc_free(mem_ctx);
4552 /* This control is sent to force the node to re-read the public addresses file
4553 and drop any addresses we should nnot longer host, and add new addresses
4554 that we are now able to host
4556 int32_t ctdb_control_reload_public_ips(struct ctdb_context *ctdb, struct ctdb_req_control_old *c, bool *async_reply)
4558 struct ctdb_reloadips_handle *h;
4559 pid_t parent = getpid();
4561 if (ctdb->reload_ips != NULL) {
4562 talloc_free(ctdb->reload_ips);
4563 ctdb->reload_ips = NULL;
4566 h = talloc(ctdb, struct ctdb_reloadips_handle);
4567 CTDB_NO_MEMORY(ctdb, h);
4572 if (pipe(h->fd) == -1) {
4573 DEBUG(DEBUG_ERR,("Failed to create pipe for ctdb_freeze_lock\n"));
4578 h->child = ctdb_fork(ctdb);
4579 if (h->child == (pid_t)-1) {
4580 DEBUG(DEBUG_ERR, ("Failed to fork a child for reloadips\n"));
4588 if (h->child == 0) {
4589 signed char res = 0;
4592 debug_extra = talloc_asprintf(NULL, "reloadips:");
4594 prctl_set_comment("ctdb_reloadips");
4595 if (switch_from_server_to_client(ctdb, "reloadips-child") != 0) {
4596 DEBUG(DEBUG_CRIT,("ERROR: Failed to switch reloadips child into client mode\n"));
4599 res = ctdb_reloadips_child(ctdb);
4601 DEBUG(DEBUG_ERR,("Failed to reload ips on local node\n"));
4605 sys_write(h->fd[1], &res, 1);
4606 /* make sure we die when our parent dies */
4607 while (ctdb_kill(ctdb, parent, 0) == 0 || errno != ESRCH) {
4613 h->c = talloc_steal(h, c);
4616 set_close_on_exec(h->fd[0]);
4618 talloc_set_destructor(h, ctdb_reloadips_destructor);
4621 h->fde = tevent_add_fd(ctdb->ev, h, h->fd[0], TEVENT_FD_READ,
4622 ctdb_reloadips_child_handler, (void *)h);
4623 tevent_fd_set_auto_close(h->fde);
4625 tevent_add_timer(ctdb->ev, h, timeval_current_ofs(120, 0),
4626 ctdb_reloadips_timeout_event, h);
4628 /* we reply later */
4629 *async_reply = true;