Copyright (C) Ronnie Sahlberg 2007
Copyright (C) Andrew Tridgell 2007
+ Copyright (C) Martin Schwenke 2011
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
struct ctdb_do_takeip_state *state =
talloc_get_type(private_data, struct ctdb_do_takeip_state);
int32_t ret;
+ TDB_DATA data;
if (status != 0) {
+ struct ctdb_node *node = ctdb->nodes[ctdb->pnn];
+
if (status == -ETIME) {
ctdb_ban_self(ctdb);
}
ctdb_addr_to_str(&state->vnn->public_address),
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;
}
return;
}
+ data.dptr = (uint8_t *)ctdb_addr_to_str(&state->vnn->public_address);
+ data.dsize = strlen((char *)data.dptr) + 1;
+ DEBUG(DEBUG_INFO,(__location__ " sending TAKE_IP for '%s'\n", data.dptr));
+
+ ctdb_daemon_send_message(ctdb, ctdb->pnn, CTDB_SRVID_TAKE_IP, data);
+
+
/* the control succeeded */
ctdb_request_control_reply(ctdb, state->c, NULL, 0, NULL);
talloc_free(state);
int ret;
struct ctdb_do_updateip_state *state;
struct ctdb_iface *old = vnn->iface;
+ char *new_name;
ctdb_vnn_unassign_iface(ctdb, vnn);
ret = ctdb_vnn_assign_iface(ctdb, vnn);
return -1;
}
- if (vnn->iface == old) {
- DEBUG(DEBUG_ERR,("update of IP %s/%u trying to "
- "assin a same interface '%s'\n",
- ctdb_addr_to_str(&vnn->public_address),
- vnn->public_netmask_bits,
- old->name));
- return -1;
+ new_name = ctdb_vnn_iface_string(vnn);
+ if (old->name != NULL && new_name != NULL && !strcmp(old->name, new_name)) {
+ /* A benign update from one interface onto itself.
+ * no need to run the eventscripts in this case, just return
+ * success.
+ */
+ ctdb_request_control_reply(ctdb, c, NULL, 0, NULL);
+ return 0;
}
state = talloc(vnn, struct ctdb_do_updateip_state);
ctdb_addr_to_str(&vnn->public_address),
vnn->public_netmask_bits,
old->name,
- ctdb_vnn_iface_string(vnn)));
+ new_name));
ret = ctdb_event_script_callback(ctdb,
state,
CTDB_EVENT_UPDATE_IP,
"%s %s %s %u",
state->old->name,
- ctdb_vnn_iface_string(vnn),
+ 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, ctdb_vnn_iface_string(vnn)));
+ old->name, new_name));
talloc_free(state);
return -1;
}
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?"
- "banning ourself\n",
+ "but we have no interface assigned, has someone manually configured it? Ignore for now.\n",
ctdb_addr_to_str(&vnn->public_address)));
- ctdb_ban_self(ctdb);
- return -1;
+ return 0;
}
- if (vnn->pnn != ctdb->pnn && have_ip) {
+ 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"
"and we are node %d, banning ourself\n",
return -1;
}
+ 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;
+ }
+
if (vnn->iface) {
if (vnn->iface->link_up) {
/* only move when the rebalance gains something */
}
if (vnn->iface == NULL) {
- DEBUG(DEBUG_CRIT,(__location__ " release_ip of IP %s is known to the kernel, "
- "but we have no interface assigned, has someone manually configured it?"
- "banning ourself\n",
+ DEBUG(DEBUG_ERR,(__location__ " release_ip 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)));
- ctdb_ban_self(ctdb);
- return -1;
+ return 0;
}
DEBUG(DEBUG_NOTICE,("Release of IP %s/%u on interface %s node:%d\n",
vnn->public_address = *addr;
vnn->public_netmask_bits = mask;
vnn->pnn = -1;
+ 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]);
talloc_free(vnn);
return -1;
}
+ if (i == 0) {
+ vnn->iface = ctdb_find_iface(ctdb, vnn->ifaces[i]);
+ }
}
DLIST_ADD(ctdb->vnn, vnn);
const char *ip)
{
struct ctdb_vnn *svnn;
+ struct ctdb_iface *cur = NULL;
bool ok;
int ret;
return -1;
}
+ /* 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;
+
ret = ctdb_vnn_assign_iface(ctdb, svnn);
if (ret != 0) {
talloc_free(svnn);
return 0;
}
-struct ctdb_public_ip_list {
- struct ctdb_public_ip_list *next;
- uint32_t pnn;
- ctdb_sock_addr addr;
-};
-
-
/* Given a physical node, return the number of
public addresses that is currently assigned to this node.
*/
static void *add_ip_callback(void *parm, void *data)
{
+ struct ctdb_public_ip_list *this_ip = parm;
+ struct ctdb_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;
}
return ip_list;
}
-/*
- make any IP alias changes for public addresses that are necessary
+/*
+ * 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.
+ * Not static, so we can easily link it into a unit test.
+ *
+ * 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?
*/
-int ctdb_takeover_run(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
+uint32_t ip_distance(ctdb_sock_addr *ip1, ctdb_sock_addr *ip2)
{
- int i, num_healthy, retries;
- struct ctdb_public_ip ip;
- struct ctdb_public_ipv4 ipv4;
- uint32_t mask;
- struct ctdb_public_ip_list *all_ips, *tmp_ip;
- int maxnode, maxnum=0, minnode, minnum=0, num;
- TDB_DATA data;
- struct timeval timeout;
- struct client_async_data *async_data;
- struct ctdb_client_control_state *state;
- TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
-
+ uint32_t ip1_k[IP_KEYLEN];
+ uint32_t *t;
+ int i;
+ uint32_t x;
- ZERO_STRUCT(ip);
+ uint32_t distance = 0;
- /* Count how many completely healthy nodes we have */
- num_healthy = 0;
- for (i=0;i<nodemap->num;i++) {
- if (!(nodemap->nodes[i].flags & (NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED))) {
- num_healthy++;
+ 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;
+ }
}
}
- if (num_healthy > 0) {
- /* We have healthy nodes, so only consider them for
- serving public addresses
- */
- mask = NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED;
- } else {
- /* We didnt have any completely healthy nodes so
- use "disabled" nodes as a fallback
- */
- mask = NODE_FLAGS_INACTIVE;
- }
+ return distance;
+}
- /* 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.
+/* 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.
+ * Not static, so we can easily link it into a unit test.
+ */
+uint32_t ip_distance_2_sum(ctdb_sock_addr *ip,
+ struct ctdb_public_ip_list *ips,
+ int pnn)
+{
+ struct ctdb_public_ip_list *t;
+ uint32_t d;
- keep the tree of ips around as ctdb->ip_tree
- */
- all_ips = create_merged_ip_list(ctdb);
+ uint32_t sum = 0;
- /* If we want deterministic ip allocations, i.e. that the ip addresses
- will always be allocated the same way for a specific set of
- available/unavailable nodes.
- */
- if (1 == ctdb->tunable.deterministic_public_ips) {
- DEBUG(DEBUG_NOTICE,("Deterministic IPs enabled. Resetting all ip allocations\n"));
- for (i=0,tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next,i++) {
- tmp_ip->pnn = i%nodemap->num;
+ for (t=ips; t != NULL; t=t->next) {
+ if (t->pnn != pnn) {
+ continue;
}
- }
-
- /* mark all public addresses with a masked node as being served by
- node -1
- */
- for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
- if (tmp_ip->pnn == -1) {
+ /* 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;
}
- if (nodemap->nodes[tmp_ip->pnn].flags & mask) {
- tmp_ip->pnn = -1;
- }
+
+ d = ip_distance(ip, &(t->addr));
+ sum += d * d; /* Cheaper than pulling in math.h :-) */
}
- /* 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) {
+ return sum;
+}
+
+/* Return the LCP2 imbalance metric for addresses currently assigned
+ to the given node.
+ * Not static, so we can easily link it into a unit test.
+ */
+uint32_t lcp2_imbalance(struct ctdb_public_ip_list * all_ips, int pnn)
+{
+ struct ctdb_public_ip_list *t;
+
+ uint32_t imbalance = 0;
+
+ for (t=all_ips; t!=NULL; t=t->next) {
+ if (t->pnn != pnn) {
continue;
}
- if (can_node_serve_ip(ctdb, tmp_ip->pnn, tmp_ip) != 0) {
- /* this node can not serve this ip. */
- tmp_ip->pnn = -1;
- }
+ /* 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.
+ * Not static, so we can easily link it into a unit test.
+ */
+void basic_allocate_unassigned(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ uint32_t mask,
+ struct ctdb_public_ip_list *all_ips)
+{
+ struct ctdb_public_ip_list *tmp_ip;
- /* now we must redistribute all public addresses with takeover node
- -1 among the nodes available
- */
- retries = 0;
-try_again:
/* loop over all ip's and find a physical node to cover for
each unassigned ip.
*/
}
}
}
+}
- /* If we dont want ips to fail back after a node becomes healthy
- again, we wont even try to reallocat the ip addresses so that
- they are evenly spread out.
- This can NOT be used at the same time as DeterministicIPs !
- */
- if (1 == ctdb->tunable.no_ip_failback) {
- if (1 == ctdb->tunable.deterministic_public_ips) {
- DEBUG(DEBUG_ERR, ("ERROR: You can not use 'DeterministicIPs' and 'NoIPFailback' at the same time\n"));
- }
- goto finished;
- }
-
+/* Basic non-deterministic rebalancing algorithm.
+ * Not static, so we can easily link it into a unit test.
+ */
+bool basic_failback(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ uint32_t mask,
+ struct ctdb_public_ip_list *all_ips,
+ int num_ips,
+ int *retries)
+{
+ int i;
+ int maxnode, maxnum=0, minnode, minnum=0, num;
+ struct ctdb_public_ip_list *tmp_ip;
- /* now, try to make sure the ip adresses are evenly distributed
- across the node.
- 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 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) {
/* 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 at most 5 times since we dont want to spend
- too much time balancing the ip coverage.
+ 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 < 5) ){
+ && (*retries < (num_ips + 5)) ){
struct ctdb_public_ip_list *tmp;
/* mark one of maxnode's vnn's as unassigned and try
for (tmp=all_ips;tmp;tmp=tmp->next) {
if (tmp->pnn == maxnode) {
tmp->pnn = -1;
- retries++;
- goto try_again;
+ (*retries)++;
+ return true;
+ }
+ }
+ }
+ }
+
+ return false;
+}
+
+/* Do necessary LCP2 initialisation. Bury it in a function here so
+ * that we can unit test it.
+ * Not static, so we can easily link it into a unit test.
+ */
+void lcp2_init(struct ctdb_context * tmp_ctx,
+ struct ctdb_node_map * nodemap,
+ uint32_t mask,
+ struct ctdb_public_ip_list *all_ips,
+ uint32_t **lcp2_imbalances,
+ bool **newly_healthy)
+{
+ int i;
+ struct ctdb_public_ip_list *tmp_ip;
+
+ *newly_healthy = talloc_array(tmp_ctx, bool, nodemap->num);
+ CTDB_NO_MEMORY_FATAL(tmp_ctx, *newly_healthy);
+ *lcp2_imbalances = talloc_array(tmp_ctx, uint32_t, nodemap->num);
+ CTDB_NO_MEMORY_FATAL(tmp_ctx, *lcp2_imbalances);
+
+ for (i=0;i<nodemap->num;i++) {
+ (*lcp2_imbalances)[i] = lcp2_imbalance(all_ips, i);
+ /* First step: is the node "healthy"? */
+ (*newly_healthy)[i] = ! (bool)(nodemap->nodes[i].flags & mask);
+ }
+
+ /* 2nd step: if a ndoe has IPs assigned then it must have been
+ * healthy before, so we remove it from consideration... */
+ for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
+ if (tmp_ip->pnn != -1) {
+ (*newly_healthy)[tmp_ip->pnn] = false;
+ }
+ }
+}
+
+/* Allocate any unassigned addresses using the LCP2 algorithm to find
+ * the IP/node combination that will cost the least.
+ * Not static, so we can easily link it into a unit test.
+ */
+void lcp2_allocate_unassigned(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ uint32_t mask,
+ struct ctdb_public_ip_list *all_ips,
+ uint32_t *lcp2_imbalances)
+{
+ struct ctdb_public_ip_list *tmp_ip;
+ int dstnode;
+
+ int minnode;
+ uint32_t mindsum, dstdsum, dstimbl, minimbl;
+ struct ctdb_public_ip_list *minip;
+
+ bool should_loop = true;
+ bool have_unassigned = true;
+
+ 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 < nodemap->num; dstnode++) {
+ /* only check nodes that can actually serve this ip */
+ if (can_node_serve_ip(ctdb, dstnode, tmp_ip)) {
+ /* no it couldnt so skip to the next node */
+ continue;
}
+ if (nodemap->nodes[dstnode].flags & mask) {
+ 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.
+ *
+ * Not static, so we can easily link it into a unit test.
+ */
+bool lcp2_failback_candidate(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ struct ctdb_public_ip_list *all_ips,
+ int srcnode,
+ uint32_t candimbl,
+ uint32_t *lcp2_imbalances,
+ bool *newly_healthy)
+{
+ int dstnode, mindstnode;
+ uint32_t srcimbl, srcdsum, dstimbl, dstdsum;
+ uint32_t minsrcimbl, mindstimbl;
+ struct ctdb_public_ip_list *minip;
+ struct ctdb_public_ip_list *tmp_ip;
+
+ /* Find an IP and destination node that best reduces imbalance. */
+ minip = NULL;
+ minsrcimbl = 0;
+ mindstnode = -1;
+ mindstimbl = 0;
+
+ DEBUG(DEBUG_DEBUG,(" ----------------------------------------\n"));
+ DEBUG(DEBUG_DEBUG,(" CONSIDERING MOVES FROM %d [%d]\n", srcnode, candimbl));
+
+ 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 = candimbl - 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 < nodemap->num; dstnode++) {
+ if (! newly_healthy[dstnode]) {
+ continue;
+ }
+ /* only check nodes that can actually serve this ip */
+ if (can_node_serve_ip(ctdb, 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, srcimbl - lcp2_imbalances[srcnode],
+ ctdb_addr_to_str(&(tmp_ip->addr)),
+ dstnode, dstimbl - lcp2_imbalances[dstnode]));
+
+ if ((dstimbl < candimbl) && (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] = srcimbl;
+ lcp2_imbalances[mindstnode] = mindstimbl;
+ minip->pnn = mindstnode;
+
+ return true;
+ }
+
+ return false;
+
+}
+
+struct lcp2_imbalance_pnn {
+ uint32_t imbalance;
+ int pnn;
+};
+
+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.
+ *
+ * Not static, so we can easily link it into a unit test.
+ */
+bool lcp2_failback(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ uint32_t mask,
+ struct ctdb_public_ip_list *all_ips,
+ uint32_t *lcp2_imbalances,
+ bool *newly_healthy)
+{
+ int i, num_newly_healthy;
+ struct lcp2_imbalance_pnn * lips;
+ bool ret;
+
+ /* It is only worth continuing if we have suitable target
+ * nodes to transfer IPs to. This check is much cheaper than
+ * continuing on...
+ */
+ num_newly_healthy = 0;
+ for (i = 0; i < nodemap->num; i++) {
+ if (newly_healthy[i]) {
+ num_newly_healthy++;
+ }
+ }
+ if (num_newly_healthy == 0) {
+ return false;
+ }
+
+ /* 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...
+ */
+ lips = talloc_array(ctdb, struct lcp2_imbalance_pnn, nodemap->num);
+ for (i = 0; i < nodemap->num; i++) {
+ lips[i].imbalance = lcp2_imbalances[i];
+ lips[i].pnn = i;
+ }
+ qsort(lips, nodemap->num, sizeof(struct lcp2_imbalance_pnn),
+ lcp2_cmp_imbalance_pnn);
+
+ ret = false;
+ for (i = 0; i < nodemap->num; 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,
+ nodemap,
+ all_ips,
+ lips[i].pnn,
+ lips[i].imbalance,
+ lcp2_imbalances,
+ newly_healthy)) {
+ ret = true;
+ break;
+ }
+ }
+
+ talloc_free(lips);
+ return ret;
+}
+
+/* The calculation part of the IP allocation algorithm.
+ * Not static, so we can easily link it into a unit test.
+ */
+void ctdb_takeover_run_core(struct ctdb_context *ctdb,
+ struct ctdb_node_map *nodemap,
+ struct ctdb_public_ip_list **all_ips_p)
+{
+ int i, num_healthy, retries, num_ips;
+ uint32_t mask;
+ struct ctdb_public_ip_list *all_ips, *tmp_ip;
+ uint32_t *lcp2_imbalances;
+ bool *newly_healthy;
+
+ TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
+
+ /* Count how many completely healthy nodes we have */
+ num_healthy = 0;
+ for (i=0;i<nodemap->num;i++) {
+ if (!(nodemap->nodes[i].flags & (NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED))) {
+ num_healthy++;
+ }
+ }
+
+ if (num_healthy > 0) {
+ /* We have healthy nodes, so only consider them for
+ serving public addresses
+ */
+ mask = NODE_FLAGS_INACTIVE|NODE_FLAGS_DISABLED;
+ } else {
+ /* We didnt have any completely healthy nodes so
+ use "disabled" nodes as a fallback
+ */
+ mask = NODE_FLAGS_INACTIVE;
+ }
+
+ /* 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);
+ *all_ips_p = all_ips; /* minimal code changes */
+
+ /* Count how many ips we have */
+ num_ips = 0;
+ for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
+ num_ips++;
+ }
+
+ /* If we want deterministic ip allocations, i.e. that the ip addresses
+ will always be allocated the same way for a specific set of
+ available/unavailable nodes.
+ */
+ if (1 == ctdb->tunable.deterministic_public_ips) {
+ DEBUG(DEBUG_NOTICE,("Deterministic IPs enabled. Resetting all ip allocations\n"));
+ for (i=0,tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next,i++) {
+ tmp_ip->pnn = i%nodemap->num;
+ }
+ }
+
+
+ /* mark all public addresses with a masked node as being served by
+ node -1
+ */
+ for (tmp_ip=all_ips;tmp_ip;tmp_ip=tmp_ip->next) {
+ if (tmp_ip->pnn == -1) {
+ continue;
+ }
+ if (nodemap->nodes[tmp_ip->pnn].flags & mask) {
+ tmp_ip->pnn = -1;
+ }
+ }
+ /* 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_serve_ip(ctdb, tmp_ip->pnn, tmp_ip) != 0) {
+ /* this node can not serve this ip. */
+ tmp_ip->pnn = -1;
+ }
+ }
+
+ if (1 == ctdb->tunable.lcp2_public_ip_assignment) {
+ lcp2_init(tmp_ctx, nodemap, mask, all_ips, &lcp2_imbalances, &newly_healthy);
+ }
+
+ /* now we must redistribute all public addresses with takeover node
+ -1 among the nodes available
+ */
+ retries = 0;
+try_again:
+ if (1 == ctdb->tunable.lcp2_public_ip_assignment) {
+ lcp2_allocate_unassigned(ctdb, nodemap, mask, all_ips, lcp2_imbalances);
+ } else {
+ basic_allocate_unassigned(ctdb, nodemap, mask, all_ips);
+ }
+
+ /* If we dont want ips to fail back after a node becomes healthy
+ again, we wont even try to reallocat the ip addresses so that
+ they are evenly spread out.
+ This can NOT be used at the same time as DeterministicIPs !
+ */
+ if (1 == ctdb->tunable.no_ip_failback) {
+ if (1 == ctdb->tunable.deterministic_public_ips) {
+ DEBUG(DEBUG_ERR, ("ERROR: You can not use 'DeterministicIPs' and 'NoIPFailback' at the same time\n"));
+ }
+ goto finished;
+ }
+
+
+ /* now, try to make sure the ip adresses are evenly distributed
+ across the node.
+ */
+ if (1 == ctdb->tunable.lcp2_public_ip_assignment) {
+ if (lcp2_failback(ctdb, nodemap, mask, all_ips, lcp2_imbalances, newly_healthy)) {
+ goto try_again;
+ }
+ } else {
+ if (basic_failback(ctdb, nodemap, mask, all_ips, num_ips, &retries)) {
+ goto try_again;
+ }
+ }
/* finished distributing the public addresses, now just send the
info out to the nodes
or -1 if there is no node that can cover this ip
*/
+ return;
+}
+
+/*
+ make any IP alias changes for public addresses that are necessary
+ */
+int ctdb_takeover_run(struct ctdb_context *ctdb, struct ctdb_node_map *nodemap)
+{
+ int i;
+ struct ctdb_public_ip ip;
+ struct ctdb_public_ipv4 ipv4;
+ uint32_t *nodes;
+ struct ctdb_public_ip_list *all_ips, *tmp_ip;
+ TDB_DATA data;
+ struct timeval timeout;
+ struct client_async_data *async_data;
+ struct ctdb_client_control_state *state;
+ TALLOC_CTX *tmp_ctx = talloc_new(ctdb);
+
+ /*
+ * ip failover is completely disabled, just send out the
+ * ipreallocated event.
+ */
+ if (ctdb->tunable.disable_ip_failover != 0) {
+ goto ipreallocated;
+ }
+
+ ZERO_STRUCT(ip);
+
+ /* Do the IP reassignment calculations */
+ ctdb_takeover_run_core(ctdb, nodemap, &all_ips);
+
/* now tell all nodes to delete any alias that they should not
have. This will be a NOOP on nodes that don't currently
hold the given alias */
return -1;
}
+ipreallocated:
+ /* tell all nodes to update natwg */
+ /* send the flags update natgw on all connected nodes */
+ data.dptr = discard_const("ipreallocated");
+ data.dsize = strlen((char *)data.dptr) + 1;
+ nodes = list_of_connected_nodes(ctdb, nodemap, tmp_ctx, true);
+ if (ctdb_client_async_control(ctdb, CTDB_CONTROL_RUN_EVENTSCRIPTS,
+ nodes, 0, TAKEOVER_TIMEOUT(),
+ false, data,
+ NULL, NULL,
+ NULL) != 0) {
+ DEBUG(DEBUG_ERR, (__location__ " ctdb_control to updatenatgw failed\n"));
+ }
+
talloc_free(tmp_ctx);
return 0;
}
struct ctdb_control_tcp_addr new_addr;
struct ctdb_control_tcp_addr *tcp_sock = NULL;
struct ctdb_tcp_list *tcp;
- struct ctdb_control_tcp_vnn t;
+ struct ctdb_tcp_connection t;
int ret;
TDB_DATA data;
struct ctdb_client_ip *ip;
DLIST_ADD(client->tcp_list, tcp);
- t.src = tcp_sock->src;
- t.dest = tcp_sock->dest;
+ t.src_addr = tcp_sock->src;
+ t.dst_addr = tcp_sock->dest;
data.dptr = (uint8_t *)&t;
data.dsize = sizeof(t);
return NULL;
}
+
+
/*
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
*/
-int32_t ctdb_control_tcp_add(struct ctdb_context *ctdb, TDB_DATA indata)
+int32_t ctdb_control_tcp_add(struct ctdb_context *ctdb, TDB_DATA indata, bool tcp_update_needed)
{
- struct ctdb_control_tcp_vnn *p = (struct ctdb_control_tcp_vnn *)indata.dptr;
+ struct ctdb_tcp_connection *p = (struct ctdb_tcp_connection *)indata.dptr;
struct ctdb_tcp_array *tcparray;
struct ctdb_tcp_connection tcp;
struct ctdb_vnn *vnn;
- vnn = find_public_ip_vnn(ctdb, &p->dest);
+ vnn = find_public_ip_vnn(ctdb, &p->dst_addr);
if (vnn == NULL) {
DEBUG(DEBUG_INFO,(__location__ " got TCP_ADD control for an address which is not a public address '%s'\n",
- ctdb_addr_to_str(&p->dest)));
+ ctdb_addr_to_str(&p->dst_addr)));
return -1;
}
tcparray->connections = talloc_size(tcparray, sizeof(struct ctdb_tcp_connection));
CTDB_NO_MEMORY(ctdb, tcparray->connections);
- tcparray->connections[tcparray->num].src_addr = p->src;
- tcparray->connections[tcparray->num].dst_addr = p->dest;
+ tcparray->connections[tcparray->num].src_addr = p->src_addr;
+ tcparray->connections[tcparray->num].dst_addr = p->dst_addr;
tcparray->num++;
+
+ if (tcp_update_needed) {
+ vnn->tcp_update_needed = true;
+ }
return 0;
}
/* Do we already have this tickle ?*/
- tcp.src_addr = p->src;
- tcp.dst_addr = p->dest;
+ tcp.src_addr = p->src_addr;
+ tcp.dst_addr = p->dst_addr;
if (ctdb_tcp_find(vnn->tcp_array, &tcp) != NULL) {
DEBUG(DEBUG_DEBUG,("Already had tickle info for %s:%u for vnn:%u\n",
ctdb_addr_to_str(&tcp.dst_addr),
CTDB_NO_MEMORY(ctdb, tcparray->connections);
vnn->tcp_array = tcparray;
- tcparray->connections[tcparray->num].src_addr = p->src;
- tcparray->connections[tcparray->num].dst_addr = p->dest;
+ tcparray->connections[tcparray->num].src_addr = p->src_addr;
+ tcparray->connections[tcparray->num].dst_addr = p->dst_addr;
tcparray->num++;
DEBUG(DEBUG_INFO,("Added tickle info for %s:%u from vnn %u\n",
ntohs(tcp.dst_addr.ip.sin_port),
vnn->pnn));
+ if (tcp_update_needed) {
+ vnn->tcp_update_needed = true;
+ }
+
return 0;
}
}
+/*
+ called by a daemon to inform us of a TCP connection that one of its
+ clients used are no longer needed in the tickle database
+ */
+int32_t ctdb_control_tcp_remove(struct ctdb_context *ctdb, TDB_DATA indata)
+{
+ struct ctdb_tcp_connection *conn = (struct ctdb_tcp_connection *)indata.dptr;
+
+ ctdb_remove_tcp_connection(ctdb, conn);
+
+ return 0;
+}
+
+
/*
called when a daemon restarts - send all tickes for all public addresses
we are serving immediately to the new node.
iface = ctdb_find_iface(ctdb, info->name);
if (iface == NULL) {
- DEBUG(DEBUG_ERR, (__location__ "iface[%s] is unknown\n",
- info->name));
return -1;
}
/* have tried too many times, just give up */
if (con->count >= 5) {
- talloc_free(con);
+ /* can't delete in traverse: reparent to delete_cons */
+ talloc_steal(param, con);
return;
}
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, NULL);
+ 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
*/
static int ctdb_killtcp_destructor(struct ctdb_kill_tcp *killtcp)
{
- if (killtcp->vnn) {
- killtcp->vnn->killtcp = NULL;
+ 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;
}
a new structure
*/
if (killtcp == NULL) {
- killtcp = talloc_zero(ctdb, struct ctdb_kill_tcp);
+ killtcp = talloc_zero(vnn, struct ctdb_kill_tcp);
CTDB_NO_MEMORY(ctdb, killtcp);
killtcp->vnn = vnn;
TALLOC_CTX *mem_ctx;
DLIST_REMOVE(ctdb->vnn, vnn);
- if (vnn->iface == NULL) {
+ if (vnn->pnn != ctdb->pnn) {
+ if (vnn->iface != NULL) {
+ ctdb_vnn_unassign_iface(ctdb, vnn);
+ }
talloc_free(vnn);
return 0;
}
+ vnn->pnn = -1;
mem_ctx = talloc_new(ctdb);
+ talloc_steal(mem_ctx, vnn);
ret = ctdb_event_script_callback(ctdb,
mem_ctx, delete_ip_callback, mem_ctx,
false,
ctdb_vnn_iface_string(vnn),
ctdb_addr_to_str(&vnn->public_address),
vnn->public_netmask_bits);
- ctdb_vnn_unassign_iface(ctdb, vnn);
- talloc_free(vnn);
+ if (vnn->iface != NULL) {
+ ctdb_vnn_unassign_iface(ctdb, vnn);
+ }
if (ret != 0) {
return -1;
}