main_loop

kill -o daemon is still running
ping local daemon
if election_timeout out
get debug_level
get relevant tunables
get runstate
get recovery lock file from the server
get nodemap
flags
if self_ban
if stop banned frozen
Retrieve capabilities from all connected nodes
validate_recovery_master --> force_election
verify ip public ip {

ips.pnn == self && dont have ip
ips.pnn != self && have ip
} tell recmaster takeover_run

down here only run with recmaster
flags right
active nodes agree we are recmaster --> force_election get vnnmap
need recovery --> do_recovery
verify all active nodes not in recover mode --> do_recovery
hold recovery lock --> do_recovery
get remote_nodemaps --> do_recovery
num_lmasters
vnnmap->size != num_lmasters --> do_recovery
nodemap node also in vnnmap --> do_recovery
all nodes have same vnnmap
if need_takeover_run do_takeover_run

<span id="force_election"></span>

force_election

election_handler

rec = self ctdb = rec->ctdb
pnn == self out

ctdb_election_win

states
longest running
biggest pnn

release recover lock file
let recmaster = that

<h1 id="do_recovery"></h1> # do_recovery > we are rec > need_recovery = true > begin > self_ban > recover_lock_file F_SETLK 为 F_WRLCK > get list of all databases dbmap > create missing local db > create missing remote db > update use same lock files > [db_recovery_parallel](#db_recovery_parallel) > [do_takeover_run](#do_takeover_run) > send message reconfigured > need_recovery = false > end > wait rerecovery_timeout

<span id='db_recovery_parallel'></span>

db_recovery_parallel

envvar CTDB_RECOVERY_HELPER
dir CTDB_HELPER_BINDIR == /usr/libexec/ctdb/
file ctdb_recovery_helper
pipe libsocket
args[0] = fd[1]
args[1] = daemon.name = CTDB_SOCKET = /var/run/ctdb/ctdb.socket
args[2] = random !=1
exec /usr/libexec/ctdb/ctdb_recovery_helper
<log-fd> <output-fd> <ctdb-socket-path> <generation>
1 1 /var/run/ctdb/ctdbd.socket 2

<span id='do_takeover_run'></span>

do_takeover_run

is_in_progress done
begin
srvid = 0 pnn = -1
list_of_connected_nodes
disable takeover_runs 60s
ctdb_takeover_run
reenable takeover_runs
ok
end

<span id='ctdb_takeover_run'></span>

ctdb_takeover_run

• 分配 ipalloc_state的内存，包括每个节点数组
• 填充 ipalloc_state的ip分配算法
• 本地填充 ipalloc_state NoIPFailback 参数--这是一个真正的集群范围的配置，只有master使用此值
• 取所有连接的节点的 NoIPTakeover 和 NoIPHostOnAllDisabled --这各动作是分开执行的，所以在单元测试过程中可以伪造
• 填充 ipalloc_state 的 NoIPTakover
• 填充 ipalloc_state 的 NoIPHost ，衍生出节点 flags 和 NoIPHostOnAllDisabled
• 检索和填充 ipalloc_state 已知和可用的IP列表
• 如果没有可用IP地址，则提前退出
• 构建列表（已知的IPs，当前指定的节点）
• 填充节点列表以强制重新平衡 - 内部结构，目前没有办法获取，只有使用LCP2算法 增加了新的IP地址的节点
• 运行IP分配算法
• 发送 RELEASE_IP 到所有节点的 取消不应有的ips
• 发送 TAKE_IP 到所有节点的 配置应有ips
• 发送 IPREALLOCATED 所有节点（向后兼容的 hack ）

ipalloc_state_init

三种算法 ipalloc_lcp2 ==> 默认
ipalloc_deterministic ==> pnn = i % numnodes ipalloc_nondeterministic ==> min以pnn=0为基准，轮询 已有ip<min的则可收ip

<span id='ipalloc_lcp2'></span>

ipalloc_lcp2

unassign_unsuitable_ips 不匹配的ip的pnn = -1
lcp2_init
lcp2_allocate_unassigned

^运算 计算出从高到低不同=distance ipv4 32 + 32 + dis + 32 = 0 ~ 128 sum = ip 到其他每个ip distance平方和 minnode || mindstdsum rebalance_candidates
lcp2_failback
均衡所有lcp2_imbalances