europelee created ZOOKEEPER-2409:
------------------------------------
Summary: zookeeper recipes lock's c implement: function
child_floor bug
Key: ZOOKEEPER-2409
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2409
Project: ZooKeeper
Issue Type: Bug
Components: c client
Affects Versions: 3.4.8
Environment: zookeeper cluster with multi servers(not standalone
mode), and multi clients connect to different zookeeper server.
Reporter: europelee
1. start zookeeper cluster with multi servers
2. multi clients connect to different zookeeper server at the same time
3. clients use lock from zookeeper recipes lock's C implement.
for example:
Client A create a node like x-025373e3a9960050-0000000067
and Client B create a node like x-015373e3a9960050-0000000068;
A is a lock owner now, then kill A, as expect, B should become owner, but in
fact B not.
Because in zoo_lock.c, function zkr_lock_operation call child_floor to
monitoring a pre node, but child_floor has bug, it caused B not check its
prenode A.
B function child_floor just simply strcmp "x-025373e3a9960050-0000000067" with
own node "x-015373e3a9960050-0000000068",
it should only strcmp "0000000067" with "0000000068", not include session info.
besides, it is better that using binary search than travelling every node for
looking for a pre node when there exists many nodes.
fix:
static char* child_floor(char **sorted_data, int len, char *element) {
char* ret = NULL;
int begin = 0;
int end = len-1;
int index = 0;
while (begin <= end) {
index = (begin+end)/2;
int iCmpRet = strcmp(strrchr(sorted_data[index], '-')+1,
strrchr(element, '-')+1);
if (iCmpRet < 0) {
begin = index + 1;
}
else {
if (iCmpRet == 0) {
if (index - 1 >= 0) {
ret = sorted_data[index-1];
}
break;
}
else {
end = index - 1;
}
}
}
return ret;
}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
