#include "common/common.h"
#include "common/logging.h"
+#include "server/ipalloc.h"
#define TAKEOVER_TIMEOUT() timeval_current_ofs(ctdb->tunable.takeover_timeout,0)
#define CTDB_ARP_INTERVAL 1
#define CTDB_ARP_REPEAT 3
-/* Flags used in IP allocation algorithms. */
-struct ctdb_ipflags {
- bool noiptakeover;
- bool noiphost;
+struct ctdb_interface {
+ struct ctdb_interface *prev, *next;
+ const char *name;
+ bool link_up;
+ uint32_t references;
};
-enum ipalloc_algorithm {
- IPALLOC_DETERMINISTIC,
- IPALLOC_NONDETERMINISTIC,
- IPALLOC_LCP2,
-};
+/* state associated with a public ip address */
+struct ctdb_vnn {
+ struct ctdb_vnn *prev, *next;
-struct ipalloc_state {
- uint32_t num;
+ struct ctdb_interface *iface;
+ const char **ifaces;
+ ctdb_sock_addr public_address;
+ uint8_t public_netmask_bits;
- /* Arrays with data for each node */
- struct ctdb_public_ip_list_old **known_public_ips;
- struct ctdb_public_ip_list_old **available_public_ips;
+ /* the node number that is serving this public address, if any.
+ If no node serves this ip it is set to -1 */
+ int32_t pnn;
- enum ipalloc_algorithm algorithm;
- uint32_t no_ip_failback;
-};
+ /* List of clients to tickle for this public address */
+ struct ctdb_tcp_array *tcp_array;
-struct ctdb_interface {
- struct ctdb_interface *prev, *next;
- const char *name;
- bool link_up;
- uint32_t references;
+ /* whether we need to update the other nodes with changes to our list
+ of connected clients */
+ bool tcp_update_needed;
+
+ /* a context to hang sending gratious arp events off */
+ TALLOC_CTX *takeover_ctx;
+
+ /* Set to true any time an update to this VNN is in flight.
+ This helps to avoid races. */
+ bool update_in_flight;
+
+ /* If CTDB_CONTROL_DEL_PUBLIC_IP is received for this IP
+ * address then this flag is set. It will be deleted in the
+ * release IP callback. */
+ bool delete_pending;
};
-static const char *ctdb_vnn_iface_string(const struct ctdb_vnn *vnn)
+static const char *iface_string(const struct ctdb_interface *iface)
{
- if (vnn->iface) {
- return vnn->iface->name;
- }
+ return (iface != NULL ? iface->name : "__none__");
+}
- return "__none__";
+static const char *ctdb_vnn_iface_string(const struct ctdb_vnn *vnn)
+{
+ return iface_string(vnn->iface);
}
static int ctdb_add_local_iface(struct ctdb_context *ctdb, const char *iface)
{
struct ctdb_interface *i;
+ if (strlen(iface) > CTDB_IFACE_SIZE) {
+ DEBUG(DEBUG_ERR, ("Interface name too long \"%s\"\n", iface));
+ return -1;
+ }
+
/* Verify that we don't have an entry for this ip yet */
for (i=ctdb->ifaces;i;i=i->next) {
if (strcmp(i->name, iface) == 0) {
/* create a new structure for this interface */
i = talloc_zero(ctdb, struct ctdb_interface);
- CTDB_NO_MEMORY_FATAL(ctdb, i);
+ if (i == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return -1;
+ }
i->name = talloc_strdup(i, iface);
- CTDB_NO_MEMORY(ctdb, i->name);
+ if (i->name == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(i);
+ return -1;
+ }
i->link_up = true;
* foolproof. One alternative is reference counting, where the logic
* is distributed and can, therefore, be broken in multiple places.
* Another alternative is to build a red-black tree of interfaces that
- * can have addresses (by walking ctdb->vnn and ctdb->single_ip_vnn
- * once) and then walking ctdb->ifaces once and deleting those not in
- * the tree. Let's go to one of those if the naive implementation
- * causes problems... :-)
+ * can have addresses (by walking ctdb->vnn once) and then walking
+ * ctdb->ifaces once and deleting those not in the tree. Let's go to
+ * one of those if the naive implementation causes problems... :-)
*/
static void ctdb_remove_orphaned_ifaces(struct ctdb_context *ctdb,
struct ctdb_vnn *vnn)
continue;
}
- /* Is the "single IP" on this interface? */
- if ((ctdb->single_ip_vnn != NULL) &&
- (ctdb->single_ip_vnn->ifaces[0] != NULL) &&
- (strcmp(i->name, ctdb->single_ip_vnn->ifaces[0]) == 0)) {
- /* Found, next interface please... */
- continue;
- }
/* Search for a vnn with this interface. */
found = false;
for (tv=ctdb->vnn; tv; tv=tv->next) {
return 0;
}
-struct takeover_callback_state {
- struct ctdb_req_control_old *c;
- ctdb_sock_addr *addr;
- struct ctdb_vnn *vnn;
-};
-
struct ctdb_do_takeip_state {
struct ctdb_req_control_old *c;
struct ctdb_vnn *vnn;
TDB_DATA data;
if (status != 0) {
- struct ctdb_node *node = ctdb->nodes[ctdb->pnn];
-
if (status == -ETIME) {
ctdb_ban_self(ctdb);
}
ctdb_vnn_iface_string(state->vnn)));
ctdb_request_control_reply(ctdb, state->c, NULL, status, NULL);
- node->flags |= NODE_FLAGS_UNHEALTHY;
talloc_free(state);
return;
}
state = talloc(vnn, struct ctdb_do_takeip_state);
CTDB_NO_MEMORY(ctdb, state);
- state->c = talloc_steal(ctdb, c);
+ state->c = NULL;
state->vnn = vnn;
vnn->update_in_flight = true;
return -1;
}
+ state->c = talloc_steal(ctdb, c);
return 0;
}
if (status == -ETIME) {
ctdb_ban_self(ctdb);
}
- DEBUG(DEBUG_ERR,(__location__ " Failed to move IP %s from interface %s to %s\n",
- ctdb_addr_to_str(&state->vnn->public_address),
- state->old->name,
- ctdb_vnn_iface_string(state->vnn)));
+ DEBUG(DEBUG_ERR,
+ ("Failed update of IP %s from interface %s to %s\n",
+ ctdb_addr_to_str(&state->vnn->public_address),
+ iface_string(state->old),
+ ctdb_vnn_iface_string(state->vnn)));
/*
* All we can do is reset the old interface
int ret;
struct ctdb_do_updateip_state *state;
struct ctdb_interface *old = vnn->iface;
+ const char *old_name = iface_string(old);
const char *new_name;
if (vnn->update_in_flight) {
ctdb_vnn_unassign_iface(ctdb, vnn);
ret = ctdb_vnn_assign_iface(ctdb, vnn);
if (ret != 0) {
- DEBUG(DEBUG_ERR,("update of IP %s/%u failed to "
- "assin a usable interface (old iface '%s')\n",
+ DEBUG(DEBUG_ERR,("Update of IP %s/%u failed to "
+ "assign a usable interface (old iface '%s')\n",
ctdb_addr_to_str(&vnn->public_address),
vnn->public_netmask_bits,
- old->name));
+ old_name));
return -1;
}
new_name = ctdb_vnn_iface_string(vnn);
- if (old->name != NULL && new_name != NULL && !strcmp(old->name, new_name)) {
+ if (old_name != NULL && new_name != NULL &&
+ strcmp(old_name, new_name) == 0) {
/* A benign update from one interface onto itself.
* no need to run the eventscripts in this case, just return
* success.
state = talloc(vnn, struct ctdb_do_updateip_state);
CTDB_NO_MEMORY(ctdb, state);
- state->c = talloc_steal(ctdb, c);
+ state->c = NULL;
state->old = old;
state->vnn = vnn;
"interface %s to %s\n",
ctdb_addr_to_str(&vnn->public_address),
vnn->public_netmask_bits,
- old->name,
+ old_name,
new_name));
ret = ctdb_event_script_callback(ctdb,
state,
CTDB_EVENT_UPDATE_IP,
"%s %s %s %u",
- state->old->name,
+ old_name,
new_name,
ctdb_addr_to_str(&vnn->public_address),
vnn->public_netmask_bits);
if (ret != 0) {
- DEBUG(DEBUG_ERR,(__location__ " Failed update IP %s from interface %s to %s\n",
- ctdb_addr_to_str(&vnn->public_address),
- old->name, new_name));
+ DEBUG(DEBUG_ERR,
+ ("Failed update IP %s from interface %s to %s\n",
+ ctdb_addr_to_str(&vnn->public_address),
+ old_name, new_name));
talloc_free(state);
return -1;
}
+ state->c = talloc_steal(ctdb, c);
return 0;
}
return -1;
}
- if (vnn->iface == NULL && vnn->pnn == -1 && have_ip && best_iface != NULL) {
- DEBUG(DEBUG_ERR,("Taking over newly created ip\n"));
- have_ip = false;
- }
-
-
- if (vnn->iface == NULL && have_ip) {
- DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
- "but we have no interface assigned, has someone manually configured it? Ignore for now.\n",
- ctdb_addr_to_str(&vnn->public_address)));
- return 0;
- }
-
if (vnn->pnn != ctdb->pnn && have_ip && vnn->pnn != -1) {
DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
"and we have it on iface[%s], but it was assigned to node %d"
}
if (vnn->pnn == -1 && have_ip) {
- vnn->pnn = ctdb->pnn;
- DEBUG(DEBUG_CRIT,(__location__ " takeoverip of IP %s is known to the kernel, "
- "and we already have it on iface[%s], update local daemon\n",
- ctdb_addr_to_str(&vnn->public_address),
- ctdb_vnn_iface_string(vnn)));
- return 0;
+ /* This will cause connections to be reset and
+ * reestablished. However, this is a very unusual
+ * situation and doing this will completely repair the
+ * inconsistency in the VNN.
+ */
+ DEBUG(DEBUG_WARNING,
+ (__location__
+ " Doing updateip for IP %s already on an interface\n",
+ ctdb_addr_to_str(&vnn->public_address)));
+ do_updateip = true;
}
if (vnn->iface) {
return 0;
}
-/*
- kill any clients that are registered with a IP that is being released
- */
-static void release_kill_clients(struct ctdb_context *ctdb, ctdb_sock_addr *addr)
-{
- struct ctdb_client_ip *ip;
-
- DEBUG(DEBUG_INFO,("release_kill_clients for ip %s\n",
- ctdb_addr_to_str(addr)));
-
- for (ip=ctdb->client_ip_list; ip; ip=ip->next) {
- ctdb_sock_addr tmp_addr;
-
- tmp_addr = ip->addr;
- DEBUG(DEBUG_INFO,("checking for client %u with IP %s\n",
- ip->client_id,
- ctdb_addr_to_str(&ip->addr)));
-
- if (ctdb_same_ip(&tmp_addr, addr)) {
- struct ctdb_client *client = reqid_find(ctdb->idr,
- ip->client_id,
- struct ctdb_client);
- DEBUG(DEBUG_INFO,("matched client %u with IP %s and pid %u\n",
- ip->client_id,
- ctdb_addr_to_str(&ip->addr),
- client->pid));
-
- if (client->pid != 0) {
- DEBUG(DEBUG_INFO,(__location__ " Killing client pid %u for IP %s on client_id %u\n",
- (unsigned)client->pid,
- ctdb_addr_to_str(addr),
- ip->client_id));
- kill(client->pid, SIGKILL);
- }
- }
- }
-}
-
static void do_delete_ip(struct ctdb_context *ctdb, struct ctdb_vnn *vnn)
{
DLIST_REMOVE(ctdb->vnn, vnn);
talloc_free(vnn);
}
+static struct ctdb_vnn *release_ip_post(struct ctdb_context *ctdb,
+ struct ctdb_vnn *vnn,
+ ctdb_sock_addr *addr)
+{
+ TDB_DATA data;
+
+ /* Send a message to all clients of this node telling them
+ * that the cluster has been reconfigured and they should
+ * close any connections on this IP address
+ */
+ data.dptr = (uint8_t *)ctdb_addr_to_str(addr);
+ data.dsize = strlen((char *)data.dptr)+1;
+ DEBUG(DEBUG_INFO, ("Sending RELEASE_IP message for %s\n", data.dptr));
+ ctdb_daemon_send_message(ctdb, ctdb->pnn, CTDB_SRVID_RELEASE_IP, data);
+
+ ctdb_vnn_unassign_iface(ctdb, vnn);
+
+ /* Process the IP if it has been marked for deletion */
+ if (vnn->delete_pending) {
+ do_delete_ip(ctdb, vnn);
+ return NULL;
+ }
+
+ return vnn;
+}
+
+struct release_ip_callback_state {
+ struct ctdb_req_control_old *c;
+ ctdb_sock_addr *addr;
+ struct ctdb_vnn *vnn;
+ uint32_t target_pnn;
+};
+
/*
called when releaseip event finishes
*/
-static void release_ip_callback(struct ctdb_context *ctdb, int status,
+static void release_ip_callback(struct ctdb_context *ctdb, int status,
void *private_data)
{
- struct takeover_callback_state *state =
- talloc_get_type(private_data, struct takeover_callback_state);
- TDB_DATA data;
+ struct release_ip_callback_state *state =
+ talloc_get_type(private_data, struct release_ip_callback_state);
if (status == -ETIME) {
ctdb_ban_self(ctdb);
}
}
- /* send a message to all clients of this node telling them
- that the cluster has been reconfigured and they should
- release any sockets on this IP */
- data.dptr = (uint8_t *)talloc_strdup(state, ctdb_addr_to_str(state->addr));
- CTDB_NO_MEMORY_VOID(ctdb, data.dptr);
- data.dsize = strlen((char *)data.dptr)+1;
-
- DEBUG(DEBUG_INFO,(__location__ " sending RELEASE_IP for '%s'\n", data.dptr));
-
- ctdb_daemon_send_message(ctdb, ctdb->pnn, CTDB_SRVID_RELEASE_IP, data);
-
- /* kill clients that have registered with this IP */
- release_kill_clients(ctdb, state->addr);
-
- ctdb_vnn_unassign_iface(ctdb, state->vnn);
-
- /* Process the IP if it has been marked for deletion */
- if (state->vnn->delete_pending) {
- do_delete_ip(ctdb, state->vnn);
- state->vnn = NULL;
- }
+ state->vnn->pnn = state->target_pnn;
+ state->vnn = release_ip_post(ctdb, state->vnn, state->addr);
/* the control succeeded */
ctdb_request_control_reply(ctdb, state->c, NULL, 0, NULL);
talloc_free(state);
}
-static int ctdb_releaseip_destructor(struct takeover_callback_state *state)
+static int ctdb_releaseip_destructor(struct release_ip_callback_state *state)
{
if (state->vnn != NULL) {
state->vnn->update_in_flight = false;
bool *async_reply)
{
int ret;
- struct takeover_callback_state *state;
+ struct release_ip_callback_state *state;
struct ctdb_public_ip *pip = (struct ctdb_public_ip *)indata.dptr;
struct ctdb_vnn *vnn;
- char *iface;
+ const char *iface;
/* update our vnn list */
vnn = find_public_ip_vnn(ctdb, &pip->addr);
ctdb_addr_to_str(&pip->addr)));
return 0;
}
- vnn->pnn = pip->pnn;
/* stop any previous arps */
talloc_free(vnn->takeover_ctx);
vnn->takeover_ctx = NULL;
- /* Some ctdb tool commands (e.g. moveip, rebalanceip) send
- * lazy multicast to drop an IP from any node that isn't the
- * intended new node. The following causes makes ctdbd ignore
- * a release for any address it doesn't host.
+ /* RELEASE_IP controls are sent to all nodes that should not
+ * be hosting a particular IP. This serves 2 purposes. The
+ * first is to help resolve any inconsistencies. If a node
+ * does unexpectly host an IP then it will be released. The
+ * 2nd is to use a "redundant release" to tell non-takeover
+ * nodes where an IP is moving to. This is how "ctdb ip" can
+ * report the (likely) location of an IP by only asking the
+ * local node. Redundant releases need to update the PNN but
+ * are otherwise ignored.
*/
if (ctdb->tunable.disable_ip_failover == 0 && ctdb->do_checkpublicip) {
if (!ctdb_sys_have_ip(&pip->addr)) {
ctdb_addr_to_str(&pip->addr),
vnn->public_netmask_bits,
ctdb_vnn_iface_string(vnn)));
+ vnn->pnn = pip->pnn;
ctdb_vnn_unassign_iface(ctdb, vnn);
return 0;
}
DEBUG(DEBUG_DEBUG,("Redundant release of IP %s/%u (ip not held)\n",
ctdb_addr_to_str(&pip->addr),
vnn->public_netmask_bits));
+ vnn->pnn = pip->pnn;
return 0;
}
}
return -1;
}
- iface = strdup(ctdb_vnn_iface_string(vnn));
+ iface = ctdb_vnn_iface_string(vnn);
DEBUG(DEBUG_NOTICE,("Release of IP %s/%u on interface %s node:%d\n",
ctdb_addr_to_str(&pip->addr),
iface,
pip->pnn));
- state = talloc(ctdb, struct takeover_callback_state);
+ state = talloc(ctdb, struct release_ip_callback_state);
if (state == NULL) {
ctdb_set_error(ctdb, "Out of memory at %s:%d",
__FILE__, __LINE__);
- free(iface);
return -1;
}
- state->c = talloc_steal(state, c);
- state->addr = talloc(state, ctdb_sock_addr);
+ state->c = NULL;
+ state->addr = talloc(state, ctdb_sock_addr);
if (state->addr == NULL) {
ctdb_set_error(ctdb, "Out of memory at %s:%d",
__FILE__, __LINE__);
- free(iface);
talloc_free(state);
return -1;
}
*state->addr = pip->addr;
+ state->target_pnn = pip->pnn;
state->vnn = vnn;
vnn->update_in_flight = true;
iface,
ctdb_addr_to_str(&pip->addr),
vnn->public_netmask_bits);
- free(iface);
if (ret != 0) {
DEBUG(DEBUG_ERR,(__location__ " Failed to release IP %s on interface %s\n",
ctdb_addr_to_str(&pip->addr),
/* tell the control that we will be reply asynchronously */
*async_reply = true;
+ state->c = talloc_steal(state, c);
return 0;
}
uint32_t num = 0;
char *tmp;
const char *iface;
- int i;
int ret;
- tmp = strdup(ifaces);
- for (iface = strtok(tmp, ","); iface; iface = strtok(NULL, ",")) {
- if (!ctdb_sys_check_iface_exists(iface)) {
- DEBUG(DEBUG_CRIT,("Interface %s does not exist. Can not add public-address : %s\n", iface, ctdb_addr_to_str(addr)));
- free(tmp);
- return -1;
- }
- }
- free(tmp);
-
- /* Verify that we don't have an entry for this ip yet */
- for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
+ /* Verify that we don't have an entry for this IP yet */
+ for (vnn = ctdb->vnn; vnn != NULL; vnn = vnn->next) {
if (ctdb_same_sockaddr(addr, &vnn->public_address)) {
- DEBUG(DEBUG_CRIT,("Same ip '%s' specified multiple times in the public address list \n",
- ctdb_addr_to_str(addr)));
+ DEBUG(DEBUG_ERR,
+ ("Duplicate public IP address '%s'\n",
+ ctdb_addr_to_str(addr)));
return -1;
- }
+ }
}
- /* create a new vnn structure for this ip address */
+ /* Create a new VNN structure for this IP address */
vnn = talloc_zero(ctdb, struct ctdb_vnn);
- CTDB_NO_MEMORY_FATAL(ctdb, vnn);
+ if (vnn == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return -1;
+ }
vnn->ifaces = talloc_array(vnn, const char *, num + 2);
+ if (vnn->ifaces == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(vnn);
+ return -1;
+ }
tmp = talloc_strdup(vnn, ifaces);
- CTDB_NO_MEMORY_FATAL(ctdb, tmp);
+ if (tmp == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(vnn);
+ return -1;
+ }
for (iface = strtok(tmp, ","); iface; iface = strtok(NULL, ",")) {
+ if (!ctdb_sys_check_iface_exists(iface)) {
+ DEBUG(DEBUG_ERR,
+ ("Unknown interface %s for public address %s\n",
+ iface, ctdb_addr_to_str(addr)));
+ talloc_free(vnn);
+ return -1;
+ }
+
+ ret = ctdb_add_local_iface(ctdb, iface);
+ if (ret != 0) {
+ DEBUG(DEBUG_ERR,
+ ("Failed to add interface '%s' "
+ "for public address %s\n",
+ iface, ctdb_addr_to_str(addr)));
+ talloc_free(vnn);
+ return -1;
+ }
+
vnn->ifaces = talloc_realloc(vnn, vnn->ifaces, const char *, num + 2);
- CTDB_NO_MEMORY_FATAL(ctdb, vnn->ifaces);
+ if (vnn->ifaces == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(vnn);
+ return -1;
+ }
vnn->ifaces[num] = talloc_strdup(vnn, iface);
- CTDB_NO_MEMORY_FATAL(ctdb, vnn->ifaces[num]);
+ if (vnn->ifaces[num] == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(vnn);
+ return -1;
+ }
num++;
}
talloc_free(tmp);
vnn->public_address = *addr;
vnn->public_netmask_bits = mask;
vnn->pnn = -1;
- if (check_address) {
- if (ctdb_sys_have_ip(addr)) {
- DEBUG(DEBUG_ERR,("We are already hosting public address '%s'. setting PNN to ourself:%d\n", ctdb_addr_to_str(addr), ctdb->pnn));
- vnn->pnn = ctdb->pnn;
- }
- }
-
- for (i=0; vnn->ifaces[i]; i++) {
- ret = ctdb_add_local_iface(ctdb, vnn->ifaces[i]);
- if (ret != 0) {
- DEBUG(DEBUG_CRIT, (__location__ " failed to add iface[%s] "
- "for public_address[%s]\n",
- vnn->ifaces[i], ctdb_addr_to_str(addr)));
- talloc_free(vnn);
- return -1;
- }
- }
DLIST_ADD(ctdb->vnn, vnn);
return 0;
}
-int ctdb_set_single_public_ip(struct ctdb_context *ctdb,
- const char *iface,
- const char *ip)
+static struct ctdb_public_ip_list *
+ctdb_fetch_remote_public_ips(struct ctdb_context *ctdb,
+ TALLOC_CTX *mem_ctx,
+ struct ctdb_node_map_old *nodemap,
+ uint32_t public_ip_flags)
{
- struct ctdb_vnn *svnn;
- struct ctdb_interface *cur = NULL;
- bool ok;
- int ret;
-
- svnn = talloc_zero(ctdb, struct ctdb_vnn);
- CTDB_NO_MEMORY(ctdb, svnn);
-
- svnn->ifaces = talloc_array(svnn, const char *, 2);
- CTDB_NO_MEMORY(ctdb, svnn->ifaces);
- svnn->ifaces[0] = talloc_strdup(svnn->ifaces, iface);
- CTDB_NO_MEMORY(ctdb, svnn->ifaces[0]);
- svnn->ifaces[1] = NULL;
-
- ok = parse_ip(ip, iface, 0, &svnn->public_address);
- if (!ok) {
- talloc_free(svnn);
- return -1;
- }
+ int j, ret;
+ struct ctdb_public_ip_list_old *ip_list;
+ struct ctdb_public_ip_list *public_ips;
- ret = ctdb_add_local_iface(ctdb, svnn->ifaces[0]);
- if (ret != 0) {
- DEBUG(DEBUG_CRIT, (__location__ " failed to add iface[%s] "
- "for single_ip[%s]\n",
- svnn->ifaces[0],
- ctdb_addr_to_str(&svnn->public_address)));
- talloc_free(svnn);
- return -1;
+ public_ips = talloc_zero_array(mem_ctx,
+ struct ctdb_public_ip_list,
+ nodemap->num);
+ if (public_ips == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return NULL;
}
- /* assume the single public ip interface is initially "good" */
- cur = ctdb_find_iface(ctdb, iface);
- if (cur == NULL) {
- DEBUG(DEBUG_CRIT,("Can not find public interface %s used by --single-public-ip", iface));
- return -1;
- }
- cur->link_up = true;
+ for (j = 0; j < nodemap->num; j++) {
+ if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
+ continue;
+ }
- ret = ctdb_vnn_assign_iface(ctdb, svnn);
- if (ret != 0) {
- talloc_free(svnn);
- return -1;
+ /* Retrieve the list of public IPs from the
+ * node. Flags says whether it is known or
+ * available. */
+ ret = ctdb_ctrl_get_public_ips_flags(
+ ctdb, TAKEOVER_TIMEOUT(), j, public_ips,
+ public_ip_flags, &ip_list);
+ if (ret != 0) {
+ DEBUG(DEBUG_ERR,
+ ("Failed to read public IPs from node: %u\n", j));
+ talloc_free(public_ips);
+ return NULL;
+ }
+ public_ips[j].num = ip_list->num;
+ if (ip_list->num == 0) {
+ talloc_free(ip_list);
+ continue;
+ }
+ public_ips[j].ip = talloc_zero_array(public_ips,
+ struct ctdb_public_ip,
+ ip_list->num);
+ if (public_ips[j].ip == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(public_ips);
+ return NULL;
+ }
+ memcpy(public_ips[j].ip, &ip_list->ips[0],
+ sizeof(struct ctdb_public_ip) * ip_list->num);
+ talloc_free(ip_list);
}
- ctdb->single_ip_vnn = svnn;
- return 0;
+ return public_ips;
}
-struct public_ip_list {
- struct public_ip_list *next;
- uint32_t pnn;
- ctdb_sock_addr addr;
+struct get_tunable_callback_data {
+ const char *tunable;
+ uint32_t *out;
+ bool fatal;
};
-/* Given a physical node, return the number of
- public addresses that is currently assigned to this node.
-*/
-static int node_ip_coverage(int32_t pnn, struct public_ip_list *ips)
-{
- int num=0;
-
- for (;ips;ips=ips->next) {
- if (ips->pnn == pnn) {
- num++;
- }
- }
- return num;
-}
-
-
-/* Can the given node host the given IP: is the public IP known to the
- * node and is NOIPHOST unset?
-*/
-static bool can_node_host_ip(struct ctdb_context *ctdb, int32_t pnn,
- struct ctdb_ipflags ipflags,
- struct public_ip_list *ip)
+static void get_tunable_callback(struct ctdb_context *ctdb, uint32_t pnn,
+ int32_t res, TDB_DATA outdata,
+ void *callback)
{
- struct ctdb_public_ip_list_old *public_ips;
- int i;
-
- if (ipflags.noiphost) {
- return false;
- }
-
- public_ips = ctdb->ipalloc_state->available_public_ips[pnn];
+ struct get_tunable_callback_data *cd =
+ (struct get_tunable_callback_data *)callback;
+ int size;
- if (public_ips == NULL) {
- return false;
+ if (res != 0) {
+ /* Already handled in fail callback */
+ return;
}
- for (i=0; i<public_ips->num; i++) {
- if (ctdb_same_ip(&ip->addr, &public_ips->ips[i].addr)) {
- /* yes, this node can serve this public ip */
- return true;
- }
+ if (outdata.dsize != sizeof(uint32_t)) {
+ DEBUG(DEBUG_ERR,("Wrong size of returned data when reading \"%s\" tunable from node %d. Expected %d bytes but received %d bytes\n",
+ cd->tunable, pnn, (int)sizeof(uint32_t),
+ (int)outdata.dsize));
+ cd->fatal = true;
+ return;
}
- return false;
-}
-
-static bool can_node_takeover_ip(struct ctdb_context *ctdb, int32_t pnn,
- struct ctdb_ipflags ipflags,
- struct public_ip_list *ip)
-{
- if (ipflags.noiptakeover) {
- return false;
+ size = talloc_array_length(cd->out);
+ if (pnn >= size) {
+ DEBUG(DEBUG_ERR,("Got %s reply from node %d but nodemap only has %d entries\n",
+ cd->tunable, pnn, size));
+ return;
}
- return can_node_host_ip(ctdb, pnn, ipflags, ip);
+
+ cd->out[pnn] = *(uint32_t *)outdata.dptr;
}
-/* search the node lists list for a node to takeover this ip.
- pick the node that currently are serving the least number of ips
- so that the ips get spread out evenly.
-*/
-static int find_takeover_node(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *ip,
- struct public_ip_list *all_ips)
+static void get_tunable_fail_callback(struct ctdb_context *ctdb, uint32_t pnn,
+ int32_t res, TDB_DATA outdata,
+ void *callback)
{
- int pnn, min=0, num;
- int i, numnodes;
-
- numnodes = talloc_array_length(ipflags);
- pnn = -1;
- for (i=0; i<numnodes; i++) {
- /* verify that this node can serve this ip */
- if (!can_node_takeover_ip(ctdb, i, ipflags[i], ip)) {
- /* no it couldnt so skip to the next node */
- continue;
- }
-
- num = node_ip_coverage(i, all_ips);
- /* was this the first node we checked ? */
- if (pnn == -1) {
- pnn = i;
- min = num;
- } else {
- if (num < min) {
- pnn = i;
- min = num;
- }
- }
- }
- if (pnn == -1) {
- DEBUG(DEBUG_WARNING,(__location__ " Could not find node to take over public address '%s'\n",
- ctdb_addr_to_str(&ip->addr)));
+ struct get_tunable_callback_data *cd =
+ (struct get_tunable_callback_data *)callback;
- return -1;
- }
-
- ip->pnn = pnn;
- return 0;
-}
-
-#define IP_KEYLEN 4
-static uint32_t *ip_key(ctdb_sock_addr *ip)
-{
- static uint32_t key[IP_KEYLEN];
-
- bzero(key, sizeof(key));
-
- switch (ip->sa.sa_family) {
- case AF_INET:
- key[3] = htonl(ip->ip.sin_addr.s_addr);
- break;
- case AF_INET6: {
- uint32_t *s6_a32 = (uint32_t *)&(ip->ip6.sin6_addr.s6_addr);
- key[0] = htonl(s6_a32[0]);
- key[1] = htonl(s6_a32[1]);
- key[2] = htonl(s6_a32[2]);
- key[3] = htonl(s6_a32[3]);
- break;
- }
- default:
- DEBUG(DEBUG_ERR, (__location__ " ERROR, unknown family passed :%u\n", ip->sa.sa_family));
- return key;
- }
-
- return key;
-}
-
-static void *add_ip_callback(void *parm, void *data)
-{
- struct public_ip_list *this_ip = parm;
- struct public_ip_list *prev_ip = data;
-
- if (prev_ip == NULL) {
- return parm;
- }
- if (this_ip->pnn == -1) {
- this_ip->pnn = prev_ip->pnn;
- }
-
- return parm;
-}
-
-static int getips_count_callback(void *param, void *data)
-{
- struct public_ip_list **ip_list = (struct public_ip_list **)param;
- struct public_ip_list *new_ip = (struct public_ip_list *)data;
-
- new_ip->next = *ip_list;
- *ip_list = new_ip;
- return 0;
-}
-
-static int verify_remote_ip_allocation(struct ctdb_context *ctdb,
- struct ctdb_public_ip_list_old *ips,
- uint32_t pnn);
-
-static int ctdb_reload_remote_public_ips(struct ctdb_context *ctdb,
- struct ipalloc_state *ipalloc_state,
- struct ctdb_node_map_old *nodemap)
-{
- int j;
- int ret;
-
- if (ipalloc_state->num != nodemap->num) {
- DEBUG(DEBUG_ERR,
- (__location__
- " ipalloc_state->num (%d) != nodemap->num (%d) invalid param\n",
- ipalloc_state->num, nodemap->num));
- return -1;
- }
-
- for (j=0; j<nodemap->num; j++) {
- if (nodemap->nodes[j].flags & NODE_FLAGS_INACTIVE) {
- continue;
- }
-
- /* Retrieve the list of known public IPs from the node */
- ret = ctdb_ctrl_get_public_ips_flags(ctdb,
- TAKEOVER_TIMEOUT(),
- j,
- ctdb->nodes,
- 0,
- &ipalloc_state->known_public_ips[j]);
- if (ret != 0) {
- DEBUG(DEBUG_ERR,
- ("Failed to read known public IPs from node: %u\n",
- j));
- return -1;
- }
-
- if (ctdb->do_checkpublicip) {
- verify_remote_ip_allocation(ctdb,
- ipalloc_state->known_public_ips[j],
- j);
- }
-
- /* Retrieve the list of available public IPs from the node */
- ret = ctdb_ctrl_get_public_ips_flags(ctdb,
- TAKEOVER_TIMEOUT(),
- j,
- ctdb->nodes,
- CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE,
- &ipalloc_state->available_public_ips[j]);
- if (ret != 0) {
- DEBUG(DEBUG_ERR,
- ("Failed to read available public IPs from node: %u\n",
- j));
- return -1;
- }
- }
-
- return 0;
-}
-
-static struct public_ip_list *
-create_merged_ip_list(struct ctdb_context *ctdb)
-{
- int i, j;
- struct public_ip_list *ip_list;
- struct ctdb_public_ip_list_old *public_ips;
-
- if (ctdb->ip_tree != NULL) {
- talloc_free(ctdb->ip_tree);
- ctdb->ip_tree = NULL;
- }
- ctdb->ip_tree = trbt_create(ctdb, 0);
-
- for (i=0;i<ctdb->num_nodes;i++) {
- public_ips = ctdb->ipalloc_state->known_public_ips[i];
-
- if (ctdb->nodes[i]->flags & NODE_FLAGS_DELETED) {
- continue;
- }
-
- /* there were no public ips for this node */
- if (public_ips == NULL) {
- continue;
- }
-
- for (j=0;j<public_ips->num;j++) {
- struct public_ip_list *tmp_ip;
-
- tmp_ip = talloc_zero(ctdb->ip_tree, struct public_ip_list);
- CTDB_NO_MEMORY_NULL(ctdb, tmp_ip);
- /* Do not use information about IP addresses hosted
- * on other nodes, it may not be accurate */
- if (public_ips->ips[j].pnn == ctdb->nodes[i]->pnn) {
- tmp_ip->pnn = public_ips->ips[j].pnn;
- } else {
- tmp_ip->pnn = -1;
- }
- tmp_ip->addr = public_ips->ips[j].addr;
- tmp_ip->next = NULL;
-
- trbt_insertarray32_callback(ctdb->ip_tree,
- IP_KEYLEN, ip_key(&public_ips->ips[j].addr),
- add_ip_callback,
- tmp_ip);
- }
- }
-
- ip_list = NULL;
- trbt_traversearray32(ctdb->ip_tree, IP_KEYLEN, getips_count_callback, &ip_list);
-
- return ip_list;
-}
-
-/*
- * This is the length of the longtest common prefix between the IPs.
- * It is calculated by XOR-ing the 2 IPs together and counting the
- * number of leading zeroes. The implementation means that all
- * addresses end up being 128 bits long.
- *
- * FIXME? Should we consider IPv4 and IPv6 separately given that the
- * 12 bytes of 0 prefix padding will hurt the algorithm if there are
- * lots of nodes and IP addresses?
- */
-static uint32_t ip_distance(ctdb_sock_addr *ip1, ctdb_sock_addr *ip2)
-{
- uint32_t ip1_k[IP_KEYLEN];
- uint32_t *t;
- int i;
- uint32_t x;
-
- uint32_t distance = 0;
-
- memcpy(ip1_k, ip_key(ip1), sizeof(ip1_k));
- t = ip_key(ip2);
- for (i=0; i<IP_KEYLEN; i++) {
- x = ip1_k[i] ^ t[i];
- if (x == 0) {
- distance += 32;
- } else {
- /* Count number of leading zeroes.
- * FIXME? This could be optimised...
- */
- while ((x & (1 << 31)) == 0) {
- x <<= 1;
- distance += 1;
- }
- }
- }
-
- return distance;
-}
-
-/* Calculate the IP distance for the given IP relative to IPs on the
- given node. The ips argument is generally the all_ips variable
- used in the main part of the algorithm.
- */
-static uint32_t ip_distance_2_sum(ctdb_sock_addr *ip,
- struct public_ip_list *ips,
- int pnn)
-{
- struct public_ip_list *t;
- uint32_t d;
-
- uint32_t sum = 0;
-
- for (t=ips; t != NULL; t=t->next) {
- if (t->pnn != pnn) {
- continue;
- }
-
- /* Optimisation: We never calculate the distance
- * between an address and itself. This allows us to
- * calculate the effect of removing an address from a
- * node by simply calculating the distance between
- * that address and all of the exitsing addresses.
- * Moreover, we assume that we're only ever dealing
- * with addresses from all_ips so we can identify an
- * address via a pointer rather than doing a more
- * expensive address comparison. */
- if (&(t->addr) == ip) {
- continue;
- }
-
- d = ip_distance(ip, &(t->addr));
- sum += d * d; /* Cheaper than pulling in math.h :-) */
- }
-
- return sum;
-}
-
-/* Return the LCP2 imbalance metric for addresses currently assigned
- to the given node.
- */
-static uint32_t lcp2_imbalance(struct public_ip_list * all_ips, int pnn)
-{
- struct public_ip_list *t;
-
- uint32_t imbalance = 0;
-
- for (t=all_ips; t!=NULL; t=t->next) {
- if (t->pnn != pnn) {
- continue;
- }
- /* Pass the rest of the IPs rather than the whole
- all_ips input list.
- */
- imbalance += ip_distance_2_sum(&(t->addr), t->next, pnn);
- }
-
- return imbalance;
-}
-
-/* Allocate any unassigned IPs just by looping through the IPs and
- * finding the best node for each.
- */
-static void basic_allocate_unassigned(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips)
-{
- struct public_ip_list *tmp_ip;
-
- /* loop over all ip's and find a physical node to cover for
- each unassigned ip.
- */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
- if (find_takeover_node(ctdb, ipflags, tmp_ip, all_ips)) {
- DEBUG(DEBUG_WARNING,("Failed to find node to cover ip %s\n",
- ctdb_addr_to_str(&tmp_ip->addr)));
- }
- }
- }
-}
-
-/* Basic non-deterministic rebalancing algorithm.
- */
-static void basic_failback(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- int num_ips)
-{
- int i, numnodes;
- int maxnode, maxnum, minnode, minnum, num, retries;
- struct public_ip_list *tmp_ip;
-
- numnodes = talloc_array_length(ipflags);
- retries = 0;
-
-try_again:
- maxnum=0;
- minnum=0;
-
- /* for each ip address, loop over all nodes that can serve
- this ip and make sure that the difference between the node
- serving the most and the node serving the least ip's are
- not greater than 1.
- */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
- continue;
- }
-
- /* Get the highest and lowest number of ips's served by any
- valid node which can serve this ip.
- */
- maxnode = -1;
- minnode = -1;
- for (i=0; i<numnodes; i++) {
- /* only check nodes that can actually serve this ip */
- if (!can_node_takeover_ip(ctdb, i, ipflags[i], tmp_ip)) {
- /* no it couldnt so skip to the next node */
- continue;
- }
-
- num = node_ip_coverage(i, all_ips);
- if (maxnode == -1) {
- maxnode = i;
- maxnum = num;
- } else {
- if (num > maxnum) {
- maxnode = i;
- maxnum = num;
- }
- }
- if (minnode == -1) {
- minnode = i;
- minnum = num;
- } else {
- if (num < minnum) {
- minnode = i;
- minnum = num;
- }
- }
- }
- if (maxnode == -1) {
- DEBUG(DEBUG_WARNING,(__location__ " Could not find maxnode. May not be able to serve ip '%s'\n",
- ctdb_addr_to_str(&tmp_ip->addr)));
-
- continue;
- }
-
- /* if the spread between the smallest and largest coverage by
- a node is >=2 we steal one of the ips from the node with
- most coverage to even things out a bit.
- try to do this a limited number of times since we dont
- want to spend too much time balancing the ip coverage.
- */
- if ( (maxnum > minnum+1)
- && (retries < (num_ips + 5)) ){
- struct public_ip_list *tmp;
-
- /* Reassign one of maxnode's VNNs */
- for (tmp=all_ips;tmp;tmp=tmp->next) {
- if (tmp->pnn == maxnode) {
- (void)find_takeover_node(ctdb, ipflags, tmp, all_ips);
- retries++;
- goto try_again;;
- }
- }
- }
- }
-}
-
-static void lcp2_init(struct ctdb_context *tmp_ctx,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- uint32_t *force_rebalance_nodes,
- uint32_t **lcp2_imbalances,
- bool **rebalance_candidates)
-{
- int i, numnodes;
- struct public_ip_list *tmp_ip;
-
- numnodes = talloc_array_length(ipflags);
-
- *rebalance_candidates = talloc_array(tmp_ctx, bool, numnodes);
- CTDB_NO_MEMORY_FATAL(tmp_ctx, *rebalance_candidates);
- *lcp2_imbalances = talloc_array(tmp_ctx, uint32_t, numnodes);
- CTDB_NO_MEMORY_FATAL(tmp_ctx, *lcp2_imbalances);
-
- for (i=0; i<numnodes; i++) {
- (*lcp2_imbalances)[i] = lcp2_imbalance(all_ips, i);
- /* First step: assume all nodes are candidates */
- (*rebalance_candidates)[i] = true;
- }
-
- /* 2nd step: if a node has IPs assigned then it must have been
- * healthy before, so we remove it from consideration. This
- * is overkill but is all we have because we don't maintain
- * state between takeover runs. An alternative would be to
- * keep state and invalidate it every time the recovery master
- * changes.
- */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn != -1) {
- (*rebalance_candidates)[tmp_ip->pnn] = false;
- }
- }
-
- /* 3rd step: if a node is forced to re-balance then
- we allow failback onto the node */
- if (force_rebalance_nodes == NULL) {
- return;
- }
- for (i = 0; i < talloc_array_length(force_rebalance_nodes); i++) {
- uint32_t pnn = force_rebalance_nodes[i];
- if (pnn >= numnodes) {
- DEBUG(DEBUG_ERR,
- (__location__ "unknown node %u\n", pnn));
- continue;
- }
-
- DEBUG(DEBUG_NOTICE,
- ("Forcing rebalancing of IPs to node %u\n", pnn));
- (*rebalance_candidates)[pnn] = true;
- }
-}
-
-/* Allocate any unassigned addresses using the LCP2 algorithm to find
- * the IP/node combination that will cost the least.
- */
-static void lcp2_allocate_unassigned(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- uint32_t *lcp2_imbalances)
-{
- struct public_ip_list *tmp_ip;
- int dstnode, numnodes;
-
- int minnode;
- uint32_t mindsum, dstdsum, dstimbl, minimbl;
- struct public_ip_list *minip;
-
- bool should_loop = true;
- bool have_unassigned = true;
-
- numnodes = talloc_array_length(ipflags);
-
- while (have_unassigned && should_loop) {
- should_loop = false;
-
- DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
- DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES (UNASSIGNED)\n"));
-
- minnode = -1;
- mindsum = 0;
- minip = NULL;
-
- /* loop over each unassigned ip. */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn != -1) {
- continue;
- }
-
- for (dstnode=0; dstnode<numnodes; dstnode++) {
- /* only check nodes that can actually takeover this ip */
- if (!can_node_takeover_ip(ctdb, dstnode,
- ipflags[dstnode],
- tmp_ip)) {
- /* no it couldnt so skip to the next node */
- continue;
- }
-
- dstdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, dstnode);
- dstimbl = lcp2_imbalances[dstnode] + dstdsum;
- DEBUG(DEBUG_DEBUG,(" %s -> %d [+%d]\n",
- ctdb_addr_to_str(&(tmp_ip->addr)),
- dstnode,
- dstimbl - lcp2_imbalances[dstnode]));
-
-
- if ((minnode == -1) || (dstdsum < mindsum)) {
- minnode = dstnode;
- minimbl = dstimbl;
- mindsum = dstdsum;
- minip = tmp_ip;
- should_loop = true;
- }
- }
- }
-
- DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
-
- /* If we found one then assign it to the given node. */
- if (minnode != -1) {
- minip->pnn = minnode;
- lcp2_imbalances[minnode] = minimbl;
- DEBUG(DEBUG_INFO,(" %s -> %d [+%d]\n",
- ctdb_addr_to_str(&(minip->addr)),
- minnode,
- mindsum));
- }
-
- /* There might be a better way but at least this is clear. */
- have_unassigned = false;
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
- have_unassigned = true;
- }
- }
- }
-
- /* We know if we have an unassigned addresses so we might as
- * well optimise.
- */
- if (have_unassigned) {
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
- DEBUG(DEBUG_WARNING,("Failed to find node to cover ip %s\n",
- ctdb_addr_to_str(&tmp_ip->addr)));
- }
- }
- }
-}
-
-/* LCP2 algorithm for rebalancing the cluster. Given a candidate node
- * to move IPs from, determines the best IP/destination node
- * combination to move from the source node.
- */
-static bool lcp2_failback_candidate(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- int srcnode,
- uint32_t *lcp2_imbalances,
- bool *rebalance_candidates)
-{
- int dstnode, mindstnode, numnodes;
- uint32_t srcimbl, srcdsum, dstimbl, dstdsum;
- uint32_t minsrcimbl, mindstimbl;
- struct public_ip_list *minip;
- struct public_ip_list *tmp_ip;
-
- /* Find an IP and destination node that best reduces imbalance. */
- srcimbl = 0;
- minip = NULL;
- minsrcimbl = 0;
- mindstnode = -1;
- mindstimbl = 0;
-
- numnodes = talloc_array_length(ipflags);
-
- DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
- DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES FROM %d [%d]\n",
- srcnode, lcp2_imbalances[srcnode]));
-
- for (tmp_ip=all_ips; tmp_ip; tmp_ip=tmp_ip->next) {
- /* Only consider addresses on srcnode. */
- if (tmp_ip->pnn != srcnode) {
- continue;
- }
-
- /* What is this IP address costing the source node? */
- srcdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, srcnode);
- srcimbl = lcp2_imbalances[srcnode] - srcdsum;
-
- /* Consider this IP address would cost each potential
- * destination node. Destination nodes are limited to
- * those that are newly healthy, since we don't want
- * to do gratuitous failover of IPs just to make minor
- * balance improvements.
- */
- for (dstnode=0; dstnode<numnodes; dstnode++) {
- if (!rebalance_candidates[dstnode]) {
- continue;
- }
-
- /* only check nodes that can actually takeover this ip */
- if (!can_node_takeover_ip(ctdb, dstnode,
- ipflags[dstnode], tmp_ip)) {
- /* no it couldnt so skip to the next node */
- continue;
- }
-
- dstdsum = ip_distance_2_sum(&(tmp_ip->addr), all_ips, dstnode);
- dstimbl = lcp2_imbalances[dstnode] + dstdsum;
- DEBUG(DEBUG_DEBUG,(" %d [%d] -> %s -> %d [+%d]\n",
- srcnode, -srcdsum,
- ctdb_addr_to_str(&(tmp_ip->addr)),
- dstnode, dstdsum));
-
- if ((dstimbl < lcp2_imbalances[srcnode]) &&
- (dstdsum < srcdsum) && \
- ((mindstnode == -1) || \
- ((srcimbl + dstimbl) < (minsrcimbl + mindstimbl)))) {
-
- minip = tmp_ip;
- minsrcimbl = srcimbl;
- mindstnode = dstnode;
- mindstimbl = dstimbl;
- }
- }
- }
- DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
-
- if (mindstnode != -1) {
- /* We found a move that makes things better... */
- DEBUG(DEBUG_INFO,("%d [%d] -> %s -> %d [+%d]\n",
- srcnode, minsrcimbl - lcp2_imbalances[srcnode],
- ctdb_addr_to_str(&(minip->addr)),
- mindstnode, mindstimbl - lcp2_imbalances[mindstnode]));
-
-
- lcp2_imbalances[srcnode] = minsrcimbl;
- lcp2_imbalances[mindstnode] = mindstimbl;
- minip->pnn = mindstnode;
-
- return true;
- }
-
- return false;
-
-}
-
-struct lcp2_imbalance_pnn {
- uint32_t imbalance;
- int pnn;
-};
-
-static int lcp2_cmp_imbalance_pnn(const void * a, const void * b)
-{
- const struct lcp2_imbalance_pnn * lipa = (const struct lcp2_imbalance_pnn *) a;
- const struct lcp2_imbalance_pnn * lipb = (const struct lcp2_imbalance_pnn *) b;
-
- if (lipa->imbalance > lipb->imbalance) {
- return -1;
- } else if (lipa->imbalance == lipb->imbalance) {
- return 0;
- } else {
- return 1;
- }
-}
-
-/* LCP2 algorithm for rebalancing the cluster. This finds the source
- * node with the highest LCP2 imbalance, and then determines the best
- * IP/destination node combination to move from the source node.
- */
-static void lcp2_failback(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- uint32_t *lcp2_imbalances,
- bool *rebalance_candidates)
-{
- int i, numnodes;
- struct lcp2_imbalance_pnn * lips;
- bool again;
-
- numnodes = talloc_array_length(ipflags);
-
-try_again:
- /* Put the imbalances and nodes into an array, sort them and
- * iterate through candidates. Usually the 1st one will be
- * used, so this doesn't cost much...
- */
- DEBUG(DEBUG_DEBUG,("+++++++++++++++++++++++++++++++++++++++++\n"));
- DEBUG(DEBUG_DEBUG,("Selecting most imbalanced node from:\n"));
- lips = talloc_array(ctdb, struct lcp2_imbalance_pnn, numnodes);
- for (i=0; i<numnodes; i++) {
- lips[i].imbalance = lcp2_imbalances[i];
- lips[i].pnn = i;
- DEBUG(DEBUG_DEBUG,(" %d [%d]\n", i, lcp2_imbalances[i]));
- }
- qsort(lips, numnodes, sizeof(struct lcp2_imbalance_pnn),
- lcp2_cmp_imbalance_pnn);
-
- again = false;
- for (i=0; i<numnodes; i++) {
- /* This means that all nodes had 0 or 1 addresses, so
- * can't be imbalanced.
- */
- if (lips[i].imbalance == 0) {
- break;
- }
-
- if (lcp2_failback_candidate(ctdb,
- ipflags,
- all_ips,
- lips[i].pnn,
- lcp2_imbalances,
- rebalance_candidates)) {
- again = true;
- break;
- }
- }
-
- talloc_free(lips);
- if (again) {
- goto try_again;
- }
-}
-
-static void unassign_unsuitable_ips(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips)
-{
- struct public_ip_list *tmp_ip;
-
- /* verify that the assigned nodes can serve that public ip
- and set it to -1 if not
- */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
- continue;
- }
- if (!can_node_host_ip(ctdb, tmp_ip->pnn,
- ipflags[tmp_ip->pnn], tmp_ip) != 0) {
- /* this node can not serve this ip. */
- DEBUG(DEBUG_DEBUG,("Unassign IP: %s from %d\n",
- ctdb_addr_to_str(&(tmp_ip->addr)),
- tmp_ip->pnn));
- tmp_ip->pnn = -1;
- }
- }
-}
-
-static void ip_alloc_deterministic_ips(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips)
-{
- struct public_ip_list *tmp_ip;
- int i, numnodes;
-
- numnodes = talloc_array_length(ipflags);
-
- DEBUG(DEBUG_NOTICE,("Deterministic IPs enabled. Resetting all ip allocations\n"));
- /* Allocate IPs to nodes in a modulo fashion so that IPs will
- * always be allocated the same way for a specific set of
- * available/unavailable nodes.
- */
-
- for (i=0,tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next,i++) {
- tmp_ip->pnn = i % numnodes;
- }
-
- /* IP failback doesn't make sense with deterministic
- * IPs, since the modulo step above implicitly fails
- * back IPs to their "home" node.
- */
- if (1 == ctdb->ipalloc_state->no_ip_failback) {
- DEBUG(DEBUG_WARNING, ("WARNING: 'NoIPFailback' set but ignored - incompatible with 'DeterministicIPs\n"));
- }
-
- unassign_unsuitable_ips(ctdb, ipflags, all_ips);
-
- basic_allocate_unassigned(ctdb, ipflags, all_ips);
-
- /* No failback here! */
-}
-
-static void ip_alloc_nondeterministic_ips(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips)
-{
- /* This should be pushed down into basic_failback. */
- struct public_ip_list *tmp_ip;
- int num_ips = 0;
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- num_ips++;
- }
-
- unassign_unsuitable_ips(ctdb, ipflags, all_ips);
-
- basic_allocate_unassigned(ctdb, ipflags, all_ips);
-
- /* If we don't want IPs to fail back then don't rebalance IPs. */
- if (1 == ctdb->ipalloc_state->no_ip_failback) {
- return;
- }
-
- /* Now, try to make sure the ip adresses are evenly distributed
- across the nodes.
- */
- basic_failback(ctdb, ipflags, all_ips, num_ips);
-}
-
-static void ip_alloc_lcp2(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- uint32_t *force_rebalance_nodes)
-{
- uint32_t *lcp2_imbalances;
- bool *rebalance_candidates;
- int numnodes, num_rebalance_candidates, i;
-
- TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
-
- unassign_unsuitable_ips(ctdb, ipflags, all_ips);
-
- lcp2_init(tmp_ctx, ipflags, all_ips,force_rebalance_nodes,
- &lcp2_imbalances, &rebalance_candidates);
-
- lcp2_allocate_unassigned(ctdb, ipflags, all_ips, lcp2_imbalances);
-
- /* If we don't want IPs to fail back then don't rebalance IPs. */
- if (1 == ctdb->ipalloc_state->no_ip_failback) {
- goto finished;
- }
-
- /* It is only worth continuing if we have suitable target
- * nodes to transfer IPs to. This check is much cheaper than
- * continuing on...
- */
- numnodes = talloc_array_length(ipflags);
- num_rebalance_candidates = 0;
- for (i=0; i<numnodes; i++) {
- if (rebalance_candidates[i]) {
- num_rebalance_candidates++;
- }
- }
- if (num_rebalance_candidates == 0) {
- goto finished;
- }
-
- /* Now, try to make sure the ip adresses are evenly distributed
- across the nodes.
- */
- lcp2_failback(ctdb, ipflags, all_ips,
- lcp2_imbalances, rebalance_candidates);
-
-finished:
- talloc_free(tmp_ctx);
-}
-
-static bool all_nodes_are_disabled(struct ctdb_node_map_old *nodemap)
-{
- int i;
-
- for (i=0;i<nodemap->num;i++) {
- if (!(nodemap->nodes[i].flags & (NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED))) {
- /* Found one completely healthy node */
- return false;
- }
- }
-
- return true;
-}
-
-/* The calculation part of the IP allocation algorithm. */
-static void ctdb_takeover_run_core(struct ctdb_context *ctdb,
- struct ctdb_ipflags *ipflags,
- struct public_ip_list *all_ips,
- uint32_t *force_rebalance_nodes)
-{
- switch (ctdb->ipalloc_state->algorithm) {
- case IPALLOC_LCP2:
- ip_alloc_lcp2(ctdb, ipflags, all_ips, force_rebalance_nodes);
- break;
- case IPALLOC_DETERMINISTIC:
- ip_alloc_deterministic_ips(ctdb, ipflags, all_ips);
- break;
- case IPALLOC_NONDETERMINISTIC:
- ip_alloc_nondeterministic_ips(ctdb, ipflags, all_ips);
- break;
- }
-
- /* at this point ->pnn is the node which will own each IP
- or -1 if there is no node that can cover this ip
- */
-
- return;
-}
-
-struct get_tunable_callback_data {
- const char *tunable;
- uint32_t *out;
- bool fatal;
-};
-
-static void get_tunable_callback(struct ctdb_context *ctdb, uint32_t pnn,
- int32_t res, TDB_DATA outdata,
- void *callback)
-{
- struct get_tunable_callback_data *cd =
- (struct get_tunable_callback_data *)callback;
- int size;
-
- if (res != 0) {
- /* Already handled in fail callback */
- return;
- }
-
- if (outdata.dsize != sizeof(uint32_t)) {
- DEBUG(DEBUG_ERR,("Wrong size of returned data when reading \"%s\" tunable from node %d. Expected %d bytes but received %d bytes\n",
- cd->tunable, pnn, (int)sizeof(uint32_t),
- (int)outdata.dsize));
- cd->fatal = true;
- return;
- }
-
- size = talloc_array_length(cd->out);
- if (pnn >= size) {
- DEBUG(DEBUG_ERR,("Got %s reply from node %d but nodemap only has %d entries\n",
- cd->tunable, pnn, size));
- return;
- }
-
-
- cd->out[pnn] = *(uint32_t *)outdata.dptr;
-}
-
-static void get_tunable_fail_callback(struct ctdb_context *ctdb, uint32_t pnn,
- int32_t res, TDB_DATA outdata,
- void *callback)
-{
- struct get_tunable_callback_data *cd =
- (struct get_tunable_callback_data *)callback;
-
- switch (res) {
- case -ETIME:
- DEBUG(DEBUG_ERR,
- ("Timed out getting tunable \"%s\" from node %d\n",
- cd->tunable, pnn));
- cd->fatal = true;
- break;
- case -EINVAL:
- case -1:
- DEBUG(DEBUG_WARNING,
- ("Tunable \"%s\" not implemented on node %d\n",
- cd->tunable, pnn));
- break;
- default:
- DEBUG(DEBUG_ERR,
- ("Unexpected error getting tunable \"%s\" from node %d\n",
- cd->tunable, pnn));
- cd->fatal = true;
+ switch (res) {
+ case -ETIME:
+ DEBUG(DEBUG_ERR,
+ ("Timed out getting tunable \"%s\" from node %d\n",
+ cd->tunable, pnn));
+ cd->fatal = true;
+ break;
+ case -EINVAL:
+ case -1:
+ DEBUG(DEBUG_WARNING,
+ ("Tunable \"%s\" not implemented on node %d\n",
+ cd->tunable, pnn));
+ break;
+ default:
+ DEBUG(DEBUG_ERR,
+ ("Unexpected error getting tunable \"%s\" from node %d\n",
+ cd->tunable, pnn));
+ cd->fatal = true;
}
}
return tvals;
}
-/* Set internal flags for IP allocation:
- * Clear ip flags
- * Set NOIPTAKOVER ip flags from per-node NoIPTakeover tunable
- * Set NOIPHOST ip flag for each INACTIVE node
- * if all nodes are disabled:
- * Set NOIPHOST ip flags from per-node NoIPHostOnAllDisabled tunable
- * else
- * Set NOIPHOST ip flags for disabled nodes
- */
-static struct ctdb_ipflags *
-set_ipflags_internal(struct ctdb_context *ctdb,
- TALLOC_CTX *tmp_ctx,
- struct ctdb_node_map_old *nodemap,
- uint32_t *tval_noiptakeover,
- uint32_t *tval_noiphostonalldisabled)
+static struct ctdb_node_map *
+ctdb_node_map_old_to_new(TALLOC_CTX *mem_ctx,
+ const struct ctdb_node_map_old *old)
{
- int i;
- struct ctdb_ipflags *ipflags;
-
- /* Clear IP flags - implicit due to talloc_zero */
- ipflags = talloc_zero_array(tmp_ctx, struct ctdb_ipflags, nodemap->num);
- CTDB_NO_MEMORY_NULL(ctdb, ipflags);
-
- for (i=0;i<nodemap->num;i++) {
- /* Can not take IPs on node with NoIPTakeover set */
- if (tval_noiptakeover[i] != 0) {
- ipflags[i].noiptakeover = true;
- }
-
- /* Can not host IPs on INACTIVE node */
- if (nodemap->nodes[i].flags & NODE_FLAGS_INACTIVE) {
- ipflags[i].noiphost = true;
- }
- }
+ struct ctdb_node_map *new;
- if (all_nodes_are_disabled(nodemap)) {
- /* If all nodes are disabled, can not host IPs on node
- * with NoIPHostOnAllDisabled set
- */
- for (i=0;i<nodemap->num;i++) {
- if (tval_noiphostonalldisabled[i] != 0) {
- ipflags[i].noiphost = true;
- }
- }
- } else {
- /* If some nodes are not disabled, then can not host
- * IPs on DISABLED node
- */
- for (i=0;i<nodemap->num;i++) {
- if (nodemap->nodes[i].flags & NODE_FLAGS_DISABLED) {
- ipflags[i].noiphost = true;
- }
- }
+ new = talloc(mem_ctx, struct ctdb_node_map);
+ if (new == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return NULL;
}
+ new->num = old->num;
+ new->node = talloc_zero_array(new,
+ struct ctdb_node_and_flags, new->num);
+ memcpy(new->node, &old->nodes[0],
+ sizeof(struct ctdb_node_and_flags) * new->num);
- return ipflags;
+ return new;
}
-static struct ctdb_ipflags *set_ipflags(struct ctdb_context *ctdb,
- TALLOC_CTX *tmp_ctx,
- struct ctdb_node_map_old *nodemap)
+
+static bool set_ipflags(struct ctdb_context *ctdb,
+ struct ipalloc_state *ipalloc_state,
+ struct ctdb_node_map_old *nodemap)
{
uint32_t *tval_noiptakeover;
uint32_t *tval_noiphostonalldisabled;
- struct ctdb_ipflags *ipflags;
-
+ struct ctdb_node_map *new;
- tval_noiptakeover = get_tunable_from_nodes(ctdb, tmp_ctx, nodemap,
+ tval_noiptakeover = get_tunable_from_nodes(ctdb, ipalloc_state, nodemap,
"NoIPTakeover", 0);
if (tval_noiptakeover == NULL) {
- return NULL;
+ return false;
}
tval_noiphostonalldisabled =
- get_tunable_from_nodes(ctdb, tmp_ctx, nodemap,
+ get_tunable_from_nodes(ctdb, ipalloc_state, nodemap,
"NoIPHostOnAllDisabled", 0);
if (tval_noiphostonalldisabled == NULL) {
/* Caller frees tmp_ctx */
- return NULL;
+ return false;
}
- ipflags = set_ipflags_internal(ctdb, tmp_ctx, nodemap,
- tval_noiptakeover,
- tval_noiphostonalldisabled);
+ new = ctdb_node_map_old_to_new(ipalloc_state, nodemap);
+ if (new == NULL) {
+ return false;
+ }
+
+ ipalloc_set_node_flags(ipalloc_state, new,
+ tval_noiptakeover,
+ tval_noiphostonalldisabled);
talloc_free(tval_noiptakeover);
talloc_free(tval_noiphostonalldisabled);
+ talloc_free(new);
- return ipflags;
+ return true;
}
-static struct ipalloc_state * ipalloc_state_init(struct ctdb_context *ctdb,
- TALLOC_CTX *mem_ctx)
+static enum ipalloc_algorithm
+determine_algorithm(const struct ctdb_tunable_list *tunables)
{
- struct ipalloc_state *ipalloc_state =
- talloc_zero(mem_ctx, struct ipalloc_state);
- if (ipalloc_state == NULL) {
- DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
- return NULL;
- }
-
- ipalloc_state->num = ctdb->num_nodes;
- ipalloc_state->known_public_ips =
- talloc_zero_array(ipalloc_state,
- struct ctdb_public_ip_list_old *,
- ipalloc_state->num);
- if (ipalloc_state->known_public_ips == NULL) {
- DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
- talloc_free(ipalloc_state);
- return NULL;
- }
- ipalloc_state->available_public_ips =
- talloc_zero_array(ipalloc_state,
- struct ctdb_public_ip_list_old *,
- ipalloc_state->num);
- if (ipalloc_state->available_public_ips == NULL) {
- DEBUG(DEBUG_ERR, (__location__ " Out of memory\n"));
- talloc_free(ipalloc_state);
- return NULL;
- }
-
- if (1 == ctdb->tunable.lcp2_public_ip_assignment) {
- ipalloc_state->algorithm = IPALLOC_LCP2;
- } else if (1 == ctdb->tunable.deterministic_public_ips) {
- ipalloc_state->algorithm = IPALLOC_DETERMINISTIC;
+ if (1 == tunables->lcp2_public_ip_assignment) {
+ return IPALLOC_LCP2;
+ } else if (1 == tunables->deterministic_public_ips) {
+ return IPALLOC_DETERMINISTIC;
} else {
- ipalloc_state->algorithm = IPALLOC_NONDETERMINISTIC;
+ return IPALLOC_NONDETERMINISTIC;
}
-
- ipalloc_state->no_ip_failback = ctdb->tunable.no_ip_failback;
-
- return ipalloc_state;
}
-struct iprealloc_callback_data {
- bool *retry_nodes;
- int retry_count;
- client_async_callback fail_callback;
- void *fail_callback_data;
- struct ctdb_node_map_old *nodemap;
+struct takeover_callback_data {
+ uint32_t num_nodes;
+ unsigned int *fail_count;
};
-static void iprealloc_fail_callback(struct ctdb_context *ctdb, uint32_t pnn,
- int32_t res, TDB_DATA outdata,
- void *callback)
+static struct takeover_callback_data *
+takeover_callback_data_init(TALLOC_CTX *mem_ctx,
+ uint32_t num_nodes)
{
- int numnodes;
- struct iprealloc_callback_data *cd =
- (struct iprealloc_callback_data *)callback;
+ static struct takeover_callback_data *takeover_data;
- numnodes = talloc_array_length(cd->retry_nodes);
- if (pnn > numnodes) {
- DEBUG(DEBUG_ERR,
- ("ipreallocated failure from node %d, "
- "but only %d nodes in nodemap\n",
- pnn, numnodes));
- return;
+ takeover_data = talloc_zero(mem_ctx, struct takeover_callback_data);
+ if (takeover_data == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ return NULL;
}
- /* Can't run the "ipreallocated" event on a INACTIVE node */
- if (cd->nodemap->nodes[pnn].flags & NODE_FLAGS_INACTIVE) {
- DEBUG(DEBUG_WARNING,
- ("ipreallocated failed on inactive node %d, ignoring\n",
- pnn));
- return;
+ takeover_data->fail_count = talloc_zero_array(takeover_data,
+ unsigned int, num_nodes);
+ if (takeover_data->fail_count == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
+ talloc_free(takeover_data);
+ return NULL;
}
- switch (res) {
- case -ETIME:
- /* If the control timed out then that's a real error,
- * so call the real fail callback
- */
- if (cd->fail_callback) {
- cd->fail_callback(ctdb, pnn, res, outdata,
- cd->fail_callback_data);
- } else {
- DEBUG(DEBUG_WARNING,
- ("iprealloc timed out but no callback registered\n"));
- }
- break;
- default:
- /* If not a timeout then either the ipreallocated
- * eventscript (or some setup) failed. This might
- * have failed because the IPREALLOCATED control isn't
- * implemented - right now there is no way of knowing
- * because the error codes are all folded down to -1.
- * Consider retrying using EVENTSCRIPT control...
- */
- DEBUG(DEBUG_WARNING,
- ("ipreallocated failure from node %d, flagging retry\n",
- pnn));
- cd->retry_nodes[pnn] = true;
- cd->retry_count++;
- }
-}
+ takeover_data->num_nodes = num_nodes;
-struct takeover_callback_data {
- bool *node_failed;
- client_async_callback fail_callback;
- void *fail_callback_data;
- struct ctdb_node_map_old *nodemap;
-};
+ return takeover_data;
+}
static void takeover_run_fail_callback(struct ctdb_context *ctdb,
uint32_t node_pnn, int32_t res,
struct takeover_callback_data *cd =
talloc_get_type_abort(callback_data,
struct takeover_callback_data);
- int i;
-
- for (i = 0; i < cd->nodemap->num; i++) {
- if (node_pnn == cd->nodemap->nodes[i].pnn) {
- break;
- }
- }
- if (i == cd->nodemap->num) {
+ if (node_pnn >= cd->num_nodes) {
DEBUG(DEBUG_ERR, (__location__ " invalid PNN %u\n", node_pnn));
return;
}
- if (!cd->node_failed[i]) {
- cd->node_failed[i] = true;
- cd->fail_callback(ctdb, node_pnn, res, outdata,
- cd->fail_callback_data);
+ if (cd->fail_count[node_pnn] == 0) {
+ DEBUG(DEBUG_ERR,
+ ("Node %u failed the takeover run\n", node_pnn));
+ }
+
+ cd->fail_count[node_pnn]++;
+}
+
+static void takeover_run_process_failures(struct ctdb_context *ctdb,
+ struct takeover_callback_data *tcd)
+{
+ unsigned int max_fails = 0;
+ uint32_t max_pnn = -1;
+ uint32_t i;
+
+ for (i = 0; i < tcd->num_nodes; i++) {
+ if (tcd->fail_count[i] > max_fails) {
+ max_pnn = i;
+ max_fails = tcd->fail_count[i];
+ }
+ }
+
+ if (max_fails > 0) {
+ int ret;
+ TDB_DATA data;
+
+ DEBUG(DEBUG_ERR,
+ ("Sending banning credits to %u with fail count %u\n",
+ max_pnn, max_fails));
+
+ data.dptr = (uint8_t *)&max_pnn;
+ data.dsize = sizeof(uint32_t);
+ ret = ctdb_client_send_message(ctdb,
+ CTDB_BROADCAST_CONNECTED,
+ CTDB_SRVID_BANNING,
+ data);
+ if (ret != 0) {
+ DEBUG(DEBUG_ERR,
+ ("Failed to set banning credits for node %u\n",
+ max_pnn));
+ }
}
}
/*
- make any IP alias changes for public addresses that are necessary
+ * Recalculate the allocation of public IPs to nodes and have the
+ * nodes host their allocated addresses.
+ *
+ * - Initialise IP allocation state. Pass:
+ + algorithm to be used;
+ + whether IP rebalancing ("failback") should be done (this uses a
+ cluster-wide configuration variable and only the value form the
+ master node is used); and
+ * + list of nodes to force rebalance (internal structure, currently
+ * no way to fetch, only used by LCP2 for nodes that have had new
+ * IP addresses added).
+ * - Set IP flags for IP allocation based on node map and tunables
+ * NoIPTakeover/NoIPHostOnAllDisabled from all connected nodes
+ * (tunable fetching done separately so values can be faked in unit
+ * testing)
+ * - Retrieve known and available IP addresses (done separately so
+ * values can be faked in unit testing)
+ * - Use ipalloc_set_public_ips() to set known and available IP
+ addresses for allocation
+ * - If cluster can't host IP addresses then early exit
+ * - Run IP allocation algorithm
+ * - Send RELEASE_IP to all nodes for IPs they should not host
+ * - Send TAKE_IP to all nodes for IPs they should host
+ * - Send IPREALLOCATED to all nodes (with backward compatibility hack)
*/
int ctdb_takeover_run(struct ctdb_context *ctdb, struct ctdb_node_map_old *nodemap,
- uint32_t *force_rebalance_nodes,
- client_async_callback fail_callback, void *callback_data)
+ uint32_t *force_rebalance_nodes)
{
- int i, j, ret;
+ int i, ret;
struct ctdb_public_ip ip;
uint32_t *nodes;
struct public_ip_list *all_ips, *tmp_ip;
struct client_async_data *async_data;
struct ctdb_client_control_state *state;
TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
- struct ctdb_ipflags *ipflags;
struct ipalloc_state *ipalloc_state;
+ struct ctdb_public_ip_list *known_ips, *available_ips;
struct takeover_callback_data *takeover_data;
- struct iprealloc_callback_data iprealloc_data;
- bool *retry_data;
- bool can_host_ips;
+
+ /* Initialise fail callback data to be used with
+ * takeover_run_fail_callback(). A failure in any of the
+ * following steps will cause an early return, so this can be
+ * reused for each of those steps without re-initialising. */
+ takeover_data = takeover_callback_data_init(tmp_ctx,
+ nodemap->num);
+ if (takeover_data == NULL) {
+ talloc_free(tmp_ctx);
+ return -1;
+ }
+
+ /* Default timeout for early jump to IPREALLOCATED. See below
+ * for explanation of 3 times... */
+ timeout = timeval_current_ofs(3 * ctdb->tunable.takeover_timeout, 0);
/*
* ip failover is completely disabled, just send out the
goto ipreallocated;
}
- ipalloc_state = ipalloc_state_init(ctdb, tmp_ctx);
+ ipalloc_state = ipalloc_state_init(tmp_ctx, ctdb->num_nodes,
+ determine_algorithm(&ctdb->tunable),
+ (ctdb->tunable.no_ip_failback != 0),
+ force_rebalance_nodes);
if (ipalloc_state == NULL) {
talloc_free(tmp_ctx);
return -1;
}
- ctdb->ipalloc_state = ipalloc_state;
- ipflags = set_ipflags(ctdb, tmp_ctx, nodemap);
- if (ipflags == NULL) {
- DEBUG(DEBUG_ERR,("Failed to set IP flags - aborting takeover run\n"));
+ if (!set_ipflags(ctdb, ipalloc_state, nodemap)) {
+ DEBUG(DEBUG_ERR,
+ ("Failed to set IP flags - aborting takeover run\n"));
talloc_free(tmp_ctx);
return -1;
}
/* Fetch known/available public IPs from each active node */
- ret = ctdb_reload_remote_public_ips(ctdb, ipalloc_state, nodemap);
- if (ret != 0) {
+ /* Fetch lists of known public IPs from all nodes */
+ known_ips = ctdb_fetch_remote_public_ips(ctdb, ipalloc_state,
+ nodemap, 0);
+ if (known_ips == NULL) {
+ DEBUG(DEBUG_ERR, ("Failed to read known public IPs\n"));
+ talloc_free(tmp_ctx);
+ return -1;
+ }
+ available_ips = ctdb_fetch_remote_public_ips(
+ ctdb, ipalloc_state, nodemap,
+ CTDB_PUBLIC_IP_FLAGS_ONLY_AVAILABLE);
+ if (available_ips == NULL) {
+ DEBUG(DEBUG_ERR, ("Failed to read available public IPs\n"));
talloc_free(tmp_ctx);
return -1;
}
- /* Short-circuit IP allocation if no node has available IPs */
- can_host_ips = false;
- for (i=0; i < ipalloc_state->num; i++) {
- if (ipalloc_state->available_public_ips[i] != NULL) {
- can_host_ips = true;
- }
+ if (! ipalloc_set_public_ips(ipalloc_state, known_ips, available_ips)) {
+ DEBUG(DEBUG_ERR, ("Failed to set public IPs\n"));
+ talloc_free(tmp_ctx);
+ return -1;
}
- if (!can_host_ips) {
+
+ if (! ipalloc_can_host_ips(ipalloc_state)) {
DEBUG(DEBUG_WARNING,("No nodes available to host public IPs yet\n"));
- return 0;
+ goto ipreallocated;
}
- /* since nodes only know about those public addresses that
- can be served by that particular node, no single node has
- a full list of all public addresses that exist in the cluster.
- Walk over all node structures and create a merged list of
- all public addresses that exist in the cluster.
-
- keep the tree of ips around as ctdb->ip_tree
- */
- all_ips = create_merged_ip_list(ctdb);
-
/* Do the IP reassignment calculations */
- ctdb_takeover_run_core(ctdb, ipflags, all_ips, force_rebalance_nodes);
+ all_ips = ipalloc(ipalloc_state);
+ if (all_ips == NULL) {
+ talloc_free(tmp_ctx);
+ return -1;
+ }
/* Now tell all nodes to release any public IPs should not
* host. This will be a NOOP on nodes that don't currently
* hold the given IP.
*/
- takeover_data = talloc_zero(tmp_ctx, struct takeover_callback_data);
- CTDB_NO_MEMORY_FATAL(ctdb, takeover_data);
-
- takeover_data->node_failed = talloc_zero_array(tmp_ctx,
- bool, nodemap->num);
- CTDB_NO_MEMORY_FATAL(ctdb, takeover_data->node_failed);
- takeover_data->fail_callback = fail_callback;
- takeover_data->fail_callback_data = callback_data;
- takeover_data->nodemap = nodemap;
-
async_data = talloc_zero(tmp_ctx, struct client_async_data);
CTDB_NO_MEMORY_FATAL(ctdb, async_data);
ZERO_STRUCT(ip); /* Avoid valgrind warnings for union */
+ /* Each of the following stages (RELEASE_IP, TAKEOVER_IP,
+ * IPREALLOCATED) notionally has a timeout of TakeoverTimeout
+ * seconds. However, RELEASE_IP can take longer due to TCP
+ * connection killing, so sometimes needs more time.
+ * Therefore, use a cumulative timeout of TakeoverTimeout * 3
+ * seconds across all 3 stages. No explicit expiry checks are
+ * needed before each stage because tevent is smart enough to
+ * fire the timeouts even if they are in the past. Initialise
+ * this here so it explicitly covers the stages we're
+ * interested in but, in particular, not the time taken by the
+ * ipalloc().
+ */
+ timeout = timeval_current_ofs(3 * ctdb->tunable.takeover_timeout, 0);
+
/* Send a RELEASE_IP to all nodes that should not be hosting
* each IP. For each IP, all but one of these will be
* redundant. However, the redundant ones are used to tell
ip.pnn = tmp_ip->pnn;
ip.addr = tmp_ip->addr;
- timeout = TAKEOVER_TIMEOUT();
data.dsize = sizeof(ip);
data.dptr = (uint8_t *)&ip;
state = ctdb_control_send(ctdb, nodemap->nodes[i].pnn,
}
}
if (ctdb_client_async_wait(ctdb, async_data) != 0) {
- DEBUG(DEBUG_ERR,(__location__ " Async control CTDB_CONTROL_RELEASE_IP failed\n"));
- talloc_free(tmp_ctx);
- return -1;
+ DEBUG(DEBUG_ERR,
+ ("Async control CTDB_CONTROL_RELEASE_IP failed\n"));
+ goto fail;
}
talloc_free(async_data);
async_data = talloc_zero(tmp_ctx, struct client_async_data);
CTDB_NO_MEMORY_FATAL(ctdb, async_data);
- async_data->fail_callback = fail_callback;
- async_data->callback_data = callback_data;
+ async_data->fail_callback = takeover_run_fail_callback;
+ async_data->callback_data = takeover_data;
for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
if (tmp_ip->pnn == -1) {
ip.pnn = tmp_ip->pnn;
ip.addr = tmp_ip->addr;
- timeout = TAKEOVER_TIMEOUT();
data.dsize = sizeof(ip);
data.dptr = (uint8_t *)&ip;
state = ctdb_control_send(ctdb, tmp_ip->pnn,
ctdb_client_async_add(async_data, state);
}
if (ctdb_client_async_wait(ctdb, async_data) != 0) {
- DEBUG(DEBUG_ERR,(__location__ " Async control CTDB_CONTROL_TAKEOVER_IP failed\n"));
- talloc_free(tmp_ctx);
- return -1;
+ DEBUG(DEBUG_ERR,
+ ("Async control CTDB_CONTROL_TAKEOVER_IP failed\n"));
+ goto fail;
}
ipreallocated:
* IPs have moved. Once upon a time this event only used to
* update natgw.
*/
- retry_data = talloc_zero_array(tmp_ctx, bool, nodemap->num);
- CTDB_NO_MEMORY_FATAL(ctdb, retry_data);
- iprealloc_data.retry_nodes = retry_data;
- iprealloc_data.retry_count = 0;
- iprealloc_data.fail_callback = fail_callback;
- iprealloc_data.fail_callback_data = callback_data;
- iprealloc_data.nodemap = nodemap;
-
nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
ret = ctdb_client_async_control(ctdb, CTDB_CONTROL_IPREALLOCATED,
- nodes, 0, TAKEOVER_TIMEOUT(),
+ nodes, 0, timeout,
false, tdb_null,
- NULL, iprealloc_fail_callback,
- &iprealloc_data);
+ NULL, takeover_run_fail_callback,
+ takeover_data);
if (ret != 0) {
- /* If the control failed then we should retry to any
- * nodes flagged by iprealloc_fail_callback using the
- * EVENTSCRIPT control. This is a best-effort at
- * backward compatiblity when running a mixed cluster
- * where some nodes have not yet been upgraded to
- * support the IPREALLOCATED control.
- */
- DEBUG(DEBUG_WARNING,
- ("Retry ipreallocated to some nodes using eventscript control\n"));
-
- nodes = talloc_array(tmp_ctx, uint32_t,
- iprealloc_data.retry_count);
- CTDB_NO_MEMORY_FATAL(ctdb, nodes);
-
- j = 0;
- for (i=0; i<nodemap->num; i++) {
- if (iprealloc_data.retry_nodes[i]) {
- nodes[j] = i;
- j++;
- }
- }
-
- data.dptr = discard_const("ipreallocated");
- data.dsize = strlen((char *)data.dptr) + 1;
- ret = ctdb_client_async_control(ctdb,
- CTDB_CONTROL_RUN_EVENTSCRIPTS,
- nodes, 0, TAKEOVER_TIMEOUT(),
- false, data,
- NULL, fail_callback,
- callback_data);
- if (ret != 0) {
- DEBUG(DEBUG_ERR, (__location__ " failed to send control to run eventscripts with \"ipreallocated\"\n"));
- }
+ DEBUG(DEBUG_ERR,
+ ("Async CTDB_CONTROL_IPREALLOCATED control failed\n"));
+ goto fail;
}
talloc_free(tmp_ctx);
return ret;
+
+fail:
+ takeover_run_process_failures(ctdb, takeover_data);
+ talloc_free(tmp_ctx);
+ return -1;
}
vnn->tcp_update_needed = true;
}
- return 0;
-}
-
-
-/*
- called by a daemon to inform us of a TCP connection that one of its
- clients managing that should tickled with an ACK when IP takeover is
- done
- */
-static void ctdb_remove_connection(struct ctdb_context *ctdb, struct ctdb_connection *conn)
+ return 0;
+}
+
+
+static void ctdb_remove_connection(struct ctdb_vnn *vnn, struct ctdb_connection *conn)
{
struct ctdb_connection *tcpp;
- struct ctdb_vnn *vnn = find_public_ip_vnn(ctdb, &conn->dst);
if (vnn == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " unable to find public address %s\n",
- ctdb_addr_to_str(&conn->dst)));
return;
}
*/
int32_t ctdb_control_tcp_remove(struct ctdb_context *ctdb, TDB_DATA indata)
{
+ struct ctdb_vnn *vnn;
struct ctdb_connection *conn = (struct ctdb_connection *)indata.dptr;
/* If we don't have public IPs, tickles are useless */
return 0;
}
- ctdb_remove_connection(ctdb, conn);
+ vnn = find_public_ip_vnn(ctdb, &conn->dst);
+ if (vnn == NULL) {
+ DEBUG(DEBUG_ERR,
+ (__location__ " unable to find public address %s\n",
+ ctdb_addr_to_str(&conn->dst)));
+ return 0;
+ }
+
+ ctdb_remove_connection(vnn, conn);
return 0;
}
void ctdb_takeover_client_destructor_hook(struct ctdb_client *client)
{
while (client->tcp_list) {
+ struct ctdb_vnn *vnn;
struct ctdb_tcp_list *tcp = client->tcp_list;
+ struct ctdb_connection *conn = &tcp->connection;
+
DLIST_REMOVE(client->tcp_list, tcp);
- ctdb_remove_connection(client->ctdb, &tcp->connection);
+
+ vnn = find_public_ip_vnn(client->ctdb,
+ &conn->dst);
+ if (vnn == NULL) {
+ DEBUG(DEBUG_ERR,
+ (__location__ " unable to find public address %s\n",
+ ctdb_addr_to_str(&conn->dst)));
+ continue;
+ }
+
+ /* If the IP address is hosted on this node then
+ * remove the connection. */
+ if (vnn->pnn == client->ctdb->pnn) {
+ ctdb_remove_connection(vnn, conn);
+ }
+
+ /* Otherwise this function has been called because the
+ * server IP address has been released to another node
+ * and the client has exited. This means that we
+ * should not delete the connection information. The
+ * takeover node processes connections too. */
}
}
void ctdb_release_all_ips(struct ctdb_context *ctdb)
{
- struct ctdb_vnn *vnn;
+ struct ctdb_vnn *vnn, *next;
int count = 0;
if (ctdb->tunable.disable_ip_failover == 1) {
return;
}
- for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
+ for (vnn = ctdb->vnn; vnn != NULL; vnn = next) {
+ /* vnn can be freed below in release_ip_post() */
+ next = vnn->next;
+
if (!ctdb_sys_have_ip(&vnn->public_address)) {
ctdb_vnn_unassign_iface(ctdb, vnn);
continue;
}
- if (!vnn->iface) {
- continue;
- }
/* Don't allow multiple releases at once. Some code,
* particularly ctdb_tickle_sentenced_connections() is
ctdb_vnn_iface_string(vnn)));
ctdb_event_script_args(ctdb, CTDB_EVENT_RELEASE_IP, "%s %s %u",
- ctdb_vnn_iface_string(vnn),
- ctdb_addr_to_str(&vnn->public_address),
- vnn->public_netmask_bits);
- release_kill_clients(ctdb, &vnn->public_address);
- ctdb_vnn_unassign_iface(ctdb, vnn);
- vnn->update_in_flight = false;
+ ctdb_vnn_iface_string(vnn),
+ ctdb_addr_to_str(&vnn->public_address),
+ vnn->public_netmask_bits);
+ /* releaseip timeouts are converted to success, so to
+ * detect failures just check if the IP address is
+ * still there...
+ */
+ if (ctdb_sys_have_ip(&vnn->public_address)) {
+ DEBUG(DEBUG_ERR,
+ (__location__
+ " IP address %s not released\n",
+ ctdb_addr_to_str(&vnn->public_address)));
+ vnn->update_in_flight = false;
+ continue;
+ }
+
+ vnn = release_ip_post(ctdb, vnn, &vnn->public_address);
+ if (vnn != NULL) {
+ vnn->update_in_flight = false;
+ }
count++;
}
addr = (ctdb_sock_addr *)indata.dptr;
vnn = find_public_ip_vnn(ctdb, addr);
- if (vnn == NULL) {
- /* if it is not a public ip it could be our 'single ip' */
- if (ctdb->single_ip_vnn) {
- if (ctdb_same_ip(&ctdb->single_ip_vnn->public_address, addr)) {
- vnn = ctdb->single_ip_vnn;
- }
- }
- }
if (vnn == NULL) {
DEBUG(DEBUG_ERR,(__location__ " Could not get public ip info, "
"'%s'not a public address\n",
if (vnn->iface == cur) {
info->active_idx = i;
}
- strncpy(info->ifaces[i].name, cur->name, sizeof(info->ifaces[i].name)-1);
+ strncpy(info->ifaces[i].name, cur->name,
+ sizeof(info->ifaces[i].name));
+ info->ifaces[i].name[sizeof(info->ifaces[i].name)-1] = '\0';
info->ifaces[i].link_state = cur->link_up;
info->ifaces[i].references = cur->references;
}
i = 0;
for (cur=ctdb->ifaces;cur;cur=cur->next) {
- strcpy(ifaces->ifaces[i].name, cur->name);
+ strncpy(ifaces->ifaces[i].name, cur->name,
+ sizeof(ifaces->ifaces[i].name));
+ ifaces->ifaces[i].name[sizeof(ifaces->ifaces[i].name)-1] = '\0';
ifaces->ifaces[i].link_state = cur->link_up;
ifaces->ifaces[i].references = cur->references;
i++;
return 0;
}
- DEBUG(iface->link_up?DEBUG_ERR:DEBUG_NOTICE,
+ DEBUG(DEBUG_ERR,
("iface[%s] has changed it's link status %s => %s\n",
iface->name,
iface->link_up?"up":"down",
}
-/*
- structure containing the listening socket and the list of tcp connections
- that the ctdb daemon is to kill
-*/
-struct ctdb_kill_tcp {
- struct ctdb_vnn *vnn;
- struct ctdb_context *ctdb;
- int capture_fd;
- struct tevent_fd *fde;
- trbt_tree_t *connections;
- void *private_data;
-};
-
-/*
- a tcp connection that is to be killed
- */
-struct ctdb_killtcp_con {
- ctdb_sock_addr src_addr;
- ctdb_sock_addr dst_addr;
- int count;
- struct ctdb_kill_tcp *killtcp;
-};
-
-/* this function is used to create a key to represent this socketpair
- in the killtcp tree.
- this key is used to insert and lookup matching socketpairs that are
- to be tickled and RST
-*/
-#define KILLTCP_KEYLEN 10
-static uint32_t *killtcp_key(ctdb_sock_addr *src, ctdb_sock_addr *dst)
-{
- static uint32_t key[KILLTCP_KEYLEN];
-
- bzero(key, sizeof(key));
-
- if (src->sa.sa_family != dst->sa.sa_family) {
- DEBUG(DEBUG_ERR, (__location__ " ERROR, different families passed :%u vs %u\n", src->sa.sa_family, dst->sa.sa_family));
- return key;
- }
-
- switch (src->sa.sa_family) {
- case AF_INET:
- key[0] = dst->ip.sin_addr.s_addr;
- key[1] = src->ip.sin_addr.s_addr;
- key[2] = dst->ip.sin_port;
- key[3] = src->ip.sin_port;
- break;
- case AF_INET6: {
- uint32_t *dst6_addr32 =
- (uint32_t *)&(dst->ip6.sin6_addr.s6_addr);
- uint32_t *src6_addr32 =
- (uint32_t *)&(src->ip6.sin6_addr.s6_addr);
- key[0] = dst6_addr32[3];
- key[1] = src6_addr32[3];
- key[2] = dst6_addr32[2];
- key[3] = src6_addr32[2];
- key[4] = dst6_addr32[1];
- key[5] = src6_addr32[1];
- key[6] = dst6_addr32[0];
- key[7] = src6_addr32[0];
- key[8] = dst->ip6.sin6_port;
- key[9] = src->ip6.sin6_port;
- break;
- }
- default:
- DEBUG(DEBUG_ERR, (__location__ " ERROR, unknown family passed :%u\n", src->sa.sa_family));
- return key;
- }
-
- return key;
-}
-
-/*
- called when we get a read event on the raw socket
- */
-static void capture_tcp_handler(struct tevent_context *ev,
- struct tevent_fd *fde,
- uint16_t flags, void *private_data)
-{
- struct ctdb_kill_tcp *killtcp = talloc_get_type(private_data, struct ctdb_kill_tcp);
- struct ctdb_killtcp_con *con;
- ctdb_sock_addr src, dst;
- uint32_t ack_seq, seq;
-
- if (!(flags & TEVENT_FD_READ)) {
- return;
- }
-
- if (ctdb_sys_read_tcp_packet(killtcp->capture_fd,
- killtcp->private_data,
- &src, &dst,
- &ack_seq, &seq) != 0) {
- /* probably a non-tcp ACK packet */
- return;
- }
-
- /* check if we have this guy in our list of connections
- to kill
- */
- con = trbt_lookuparray32(killtcp->connections,
- KILLTCP_KEYLEN, killtcp_key(&src, &dst));
- if (con == NULL) {
- /* no this was some other packet we can just ignore */
- return;
- }
-
- /* This one has been tickled !
- now reset him and remove him from the list.
- */
- DEBUG(DEBUG_INFO, ("sending a tcp reset to kill connection :%d -> %s:%d\n",
- ntohs(con->dst_addr.ip.sin_port),
- ctdb_addr_to_str(&con->src_addr),
- ntohs(con->src_addr.ip.sin_port)));
-
- ctdb_sys_send_tcp(&con->dst_addr, &con->src_addr, ack_seq, seq, 1);
- talloc_free(con);
-}
-
-
-/* when traversing the list of all tcp connections to send tickle acks to
- (so that we can capture the ack coming back and kill the connection
- by a RST)
- this callback is called for each connection we are currently trying to kill
-*/
-static int tickle_connection_traverse(void *param, void *data)
-{
- struct ctdb_killtcp_con *con = talloc_get_type(data, struct ctdb_killtcp_con);
-
- /* have tried too many times, just give up */
- if (con->count >= 5) {
- /* can't delete in traverse: reparent to delete_cons */
- talloc_steal(param, con);
- return 0;
- }
-
- /* othervise, try tickling it again */
- con->count++;
- ctdb_sys_send_tcp(
- (ctdb_sock_addr *)&con->dst_addr,
- (ctdb_sock_addr *)&con->src_addr,
- 0, 0, 0);
- return 0;
-}
-
-
-/*
- called every second until all sentenced connections have been reset
- */
-static void ctdb_tickle_sentenced_connections(struct tevent_context *ev,
- struct tevent_timer *te,
- struct timeval t, void *private_data)
-{
- struct ctdb_kill_tcp *killtcp = talloc_get_type(private_data, struct ctdb_kill_tcp);
- void *delete_cons = talloc_new(NULL);
-
- /* loop over all connections sending tickle ACKs */
- trbt_traversearray32(killtcp->connections, KILLTCP_KEYLEN, tickle_connection_traverse, delete_cons);
-
- /* now we've finished traverse, it's safe to do deletion. */
- talloc_free(delete_cons);
-
- /* If there are no more connections to kill we can remove the
- entire killtcp structure
- */
- if ( (killtcp->connections == NULL) ||
- (killtcp->connections->root == NULL) ) {
- talloc_free(killtcp);
- return;
- }
-
- /* try tickling them again in a seconds time
- */
- tevent_add_timer(killtcp->ctdb->ev, killtcp,
- timeval_current_ofs(1, 0),
- ctdb_tickle_sentenced_connections, killtcp);
-}
-
-/*
- destroy the killtcp structure
- */
-static int ctdb_killtcp_destructor(struct ctdb_kill_tcp *killtcp)
-{
- struct ctdb_vnn *tmpvnn;
-
- /* verify that this vnn is still active */
- for (tmpvnn = killtcp->ctdb->vnn; tmpvnn; tmpvnn = tmpvnn->next) {
- if (tmpvnn == killtcp->vnn) {
- break;
- }
- }
-
- if (tmpvnn == NULL) {
- return 0;
- }
-
- if (killtcp->vnn->killtcp != killtcp) {
- return 0;
- }
-
- killtcp->vnn->killtcp = NULL;
-
- return 0;
-}
-
-
-/* nothing fancy here, just unconditionally replace any existing
- connection structure with the new one.
-
- don't even free the old one if it did exist, that one is talloc_stolen
- by the same node in the tree anyway and will be deleted when the new data
- is deleted
-*/
-static void *add_killtcp_callback(void *parm, void *data)
-{
- return parm;
-}
-
-/*
- add a tcp socket to the list of connections we want to RST
- */
-static int ctdb_killtcp_add_connection(struct ctdb_context *ctdb,
- ctdb_sock_addr *s,
- ctdb_sock_addr *d)
-{
- ctdb_sock_addr src, dst;
- struct ctdb_kill_tcp *killtcp;
- struct ctdb_killtcp_con *con;
- struct ctdb_vnn *vnn;
-
- ctdb_canonicalize_ip(s, &src);
- ctdb_canonicalize_ip(d, &dst);
-
- vnn = find_public_ip_vnn(ctdb, &dst);
- if (vnn == NULL) {
- vnn = find_public_ip_vnn(ctdb, &src);
- }
- if (vnn == NULL) {
- /* if it is not a public ip it could be our 'single ip' */
- if (ctdb->single_ip_vnn) {
- if (ctdb_same_ip(&ctdb->single_ip_vnn->public_address, &dst)) {
- vnn = ctdb->single_ip_vnn;
- }
- }
- }
- if (vnn == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " Could not killtcp, not a public address\n"));
- return -1;
- }
-
- killtcp = vnn->killtcp;
-
- /* If this is the first connection to kill we must allocate
- a new structure
- */
- if (killtcp == NULL) {
- killtcp = talloc_zero(vnn, struct ctdb_kill_tcp);
- CTDB_NO_MEMORY(ctdb, killtcp);
-
- killtcp->vnn = vnn;
- killtcp->ctdb = ctdb;
- killtcp->capture_fd = -1;
- killtcp->connections = trbt_create(killtcp, 0);
-
- vnn->killtcp = killtcp;
- talloc_set_destructor(killtcp, ctdb_killtcp_destructor);
- }
-
-
-
- /* create a structure that describes this connection we want to
- RST and store it in killtcp->connections
- */
- con = talloc(killtcp, struct ctdb_killtcp_con);
- CTDB_NO_MEMORY(ctdb, con);
- con->src_addr = src;
- con->dst_addr = dst;
- con->count = 0;
- con->killtcp = killtcp;
-
-
- trbt_insertarray32_callback(killtcp->connections,
- KILLTCP_KEYLEN, killtcp_key(&con->dst_addr, &con->src_addr),
- add_killtcp_callback, con);
-
- /*
- If we don't have a socket to listen on yet we must create it
- */
- if (killtcp->capture_fd == -1) {
- const char *iface = ctdb_vnn_iface_string(vnn);
- killtcp->capture_fd = ctdb_sys_open_capture_socket(iface, &killtcp->private_data);
- if (killtcp->capture_fd == -1) {
- DEBUG(DEBUG_CRIT,(__location__ " Failed to open capturing "
- "socket on iface '%s' for killtcp (%s)\n",
- iface, strerror(errno)));
- goto failed;
- }
- }
-
-
- if (killtcp->fde == NULL) {
- killtcp->fde = tevent_add_fd(ctdb->ev, killtcp,
- killtcp->capture_fd,
- TEVENT_FD_READ,
- capture_tcp_handler, killtcp);
- tevent_fd_set_auto_close(killtcp->fde);
-
- /* We also need to set up some events to tickle all these connections
- until they are all reset
- */
- tevent_add_timer(ctdb->ev, killtcp, timeval_current_ofs(1, 0),
- ctdb_tickle_sentenced_connections, killtcp);
- }
-
- /* tickle him once now */
- ctdb_sys_send_tcp(
- &con->dst_addr,
- &con->src_addr,
- 0, 0, 0);
-
- return 0;
-
-failed:
- talloc_free(vnn->killtcp);
- vnn->killtcp = NULL;
- return -1;
-}
-
-/*
- kill a TCP connection.
- */
-int32_t ctdb_control_kill_tcp(struct ctdb_context *ctdb, TDB_DATA indata)
-{
- struct ctdb_connection *killtcp = (struct ctdb_connection *)indata.dptr;
-
- return ctdb_killtcp_add_connection(ctdb, &killtcp->src, &killtcp->dst);
-}
-
/*
called by a daemon to inform us of the entire list of TCP tickles for
a particular public address.
return 1;
}
+ if (vnn->pnn == ctdb->pnn) {
+ DEBUG(DEBUG_INFO,
+ ("Ignoring redundant set tcp tickle list, this node hosts '%s'\n",
+ ctdb_addr_to_str(&list->addr)));
+ return 0;
+ }
+
/* remove any old ticklelist we might have */
talloc_free(vnn->tcp_array);
vnn->tcp_array = NULL;
ctdb_sock_addr *addr = (ctdb_sock_addr *)indata.dptr;
struct ctdb_tickle_list_old *list;
struct ctdb_tcp_array *tcparray;
- int num;
+ int num, i;
struct ctdb_vnn *vnn;
+ unsigned port;
vnn = find_public_ip_vnn(ctdb, addr);
if (vnn == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " Could not get tcp tickle list, '%s' is not a public address\n",
+ DEBUG(DEBUG_ERR,(__location__ " Could not get tcp tickle list, '%s' is not a public address\n",
ctdb_addr_to_str(addr)));
return 1;
}
+ port = ctdb_addr_to_port(addr);
+
tcparray = vnn->tcp_array;
- if (tcparray) {
- num = tcparray->num;
- } else {
- num = 0;
+ num = 0;
+ if (tcparray != NULL) {
+ if (port == 0) {
+ /* All connections */
+ num = tcparray->num;
+ } else {
+ /* Count connections for port */
+ for (i = 0; i < tcparray->num; i++) {
+ if (port == ctdb_addr_to_port(&tcparray->connections[i].dst)) {
+ num++;
+ }
+ }
+ }
}
outdata->dsize = offsetof(struct ctdb_tickle_list_old, connections)
list->addr = *addr;
list->num = num;
- if (num) {
- memcpy(&list->connections[0], tcparray->connections,
- sizeof(struct ctdb_connection) * num);
+
+ if (num == 0) {
+ return 0;
+ }
+
+ num = 0;
+ for (i = 0; i < tcparray->num; i++) {
+ if (port == 0 || \
+ port == ctdb_addr_to_port(&tcparray->connections[i].dst)) {
+ list->connections[num] = tcparray->connections[i];
+ num++;
+ }
}
return 0;
return 0;
}
-struct delete_ip_callback_state {
- struct ctdb_req_control_old *c;
-};
-
-/*
- called when releaseip event finishes for del_public_address
- */
-static void delete_ip_callback(struct ctdb_context *ctdb,
- int32_t status, TDB_DATA data,
- const char *errormsg,
- void *private_data)
-{
- struct delete_ip_callback_state *state =
- talloc_get_type(private_data, struct delete_ip_callback_state);
-
- /* If release failed then fail. */
- ctdb_request_control_reply(ctdb, state->c, NULL, status, errormsg);
- talloc_free(private_data);
-}
-
-int32_t ctdb_control_del_public_address(struct ctdb_context *ctdb,
- struct ctdb_req_control_old *c,
- TDB_DATA indata, bool *async_reply)
+int32_t ctdb_control_del_public_address(struct ctdb_context *ctdb, TDB_DATA indata)
{
struct ctdb_addr_info_old *pub = (struct ctdb_addr_info_old *)indata.dptr;
struct ctdb_vnn *vnn;
for (vnn=ctdb->vnn;vnn;vnn=vnn->next) {
if (ctdb_same_ip(&vnn->public_address, &pub->addr)) {
if (vnn->pnn == ctdb->pnn) {
- struct delete_ip_callback_state *state;
- struct ctdb_public_ip *ip;
- TDB_DATA data;
- int ret;
-
+ /* This IP is currently being hosted.
+ * Defer the deletion until the next
+ * takeover run. "ctdb reloadips" will
+ * always cause a takeover run. "ctdb
+ * delip" will now need an explicit
+ * "ctdb ipreallocated" afterwards. */
vnn->delete_pending = true;
-
- state = talloc(ctdb,
- struct delete_ip_callback_state);
- CTDB_NO_MEMORY(ctdb, state);
- state->c = c;
-
- ip = talloc(state, struct ctdb_public_ip);
- if (ip == NULL) {
- DEBUG(DEBUG_ERR,
- (__location__ " Out of memory\n"));
- talloc_free(state);
- return -1;
- }
- ip->pnn = -1;
- ip->addr = pub->addr;
-
- data.dsize = sizeof(struct ctdb_public_ip);
- data.dptr = (unsigned char *)ip;
-
- ret = ctdb_daemon_send_control(ctdb,
- ctdb_get_pnn(ctdb),
- 0,
- CTDB_CONTROL_RELEASE_IP,
- 0, 0,
- data,
- delete_ip_callback,
- state);
- if (ret == -1) {
- DEBUG(DEBUG_ERR,
- (__location__ "Unable to send "
- "CTDB_CONTROL_RELEASE_IP\n"));
- talloc_free(state);
- return -1;
- }
-
- state->c = talloc_steal(state, c);
- *async_reply = true;
} else {
/* This IP is not hosted on the
* current node so just delete it
}
-/* This function is called from the recovery daemon to verify that a remote
- node has the expected ip allocation.
- This is verified against ctdb->ip_tree
-*/
-static int verify_remote_ip_allocation(struct ctdb_context *ctdb,
- struct ctdb_public_ip_list_old *ips,
- uint32_t pnn)
-{
- struct public_ip_list *tmp_ip;
- int i;
-
- if (ctdb->ip_tree == NULL) {
- /* don't know the expected allocation yet, assume remote node
- is correct. */
- return 0;
- }
-
- if (ips == NULL) {
- return 0;
- }
-
- for (i=0; i<ips->num; i++) {
- tmp_ip = trbt_lookuparray32(ctdb->ip_tree, IP_KEYLEN, ip_key(&ips->ips[i].addr));
- if (tmp_ip == NULL) {
- DEBUG(DEBUG_ERR,("Node %u has new or unknown public IP %s\n", pnn, ctdb_addr_to_str(&ips->ips[i].addr)));
- return -1;
- }
-
- if (tmp_ip->pnn == -1 || ips->ips[i].pnn == -1) {
- continue;
- }
-
- if (tmp_ip->pnn != ips->ips[i].pnn) {
- DEBUG(DEBUG_ERR,
- ("Inconsistent IP allocation - node %u thinks %s is held by node %u while it is assigned to node %u\n",
- pnn,
- ctdb_addr_to_str(&ips->ips[i].addr),
- ips->ips[i].pnn, tmp_ip->pnn));
- return -1;
- }
- }
-
- return 0;
-}
-
-int update_ip_assignment_tree(struct ctdb_context *ctdb, struct ctdb_public_ip *ip)
-{
- struct public_ip_list *tmp_ip;
-
- /* IP tree is never built if DisableIPFailover is set */
- if (ctdb->tunable.disable_ip_failover != 0) {
- return 0;
- }
-
- if (ctdb->ip_tree == NULL) {
- DEBUG(DEBUG_ERR,("No ctdb->ip_tree yet. Failed to update ip assignment\n"));
- return -1;
- }
-
- tmp_ip = trbt_lookuparray32(ctdb->ip_tree, IP_KEYLEN, ip_key(&ip->addr));
- if (tmp_ip == NULL) {
- DEBUG(DEBUG_ERR,(__location__ " Could not find record for address %s, update ip\n", ctdb_addr_to_str(&ip->addr)));
- return -1;
- }
-
- 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));
- tmp_ip->pnn = ip->pnn;
-
- return 0;
-}
-
-void clear_ip_assignment_tree(struct ctdb_context *ctdb)
-{
- TALLOC_FREE(ctdb->ip_tree);
-}
-
struct ctdb_reloadips_handle {
struct ctdb_context *ctdb;
struct ctdb_req_control_old *c;
}
sys_write(h->fd[1], &res, 1);
- /* make sure we die when our parent dies */
- while (ctdb_kill(ctdb, parent, 0) == 0 || errno != ESRCH) {
- sleep(5);
- }
+ ctdb_wait_for_process_to_exit(parent);
_exit(0);
}