[jira] [Updated] (ZOOKEEPER-1446) C API makes it difficult to implement a timed wait_until_connected method correctly

["Stephen Tyree (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/ZOOKEEPER-1446?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Stephen Tyree updated ZOOKEEPER-1446:
-------------------------------------

    Attachment: ZOOKEEPER-1446.patch

An example patch. Was unable to run the unit test I wrote (because my setup is being a jerk), but I'm more looking for input on the design.
                
> C API makes it difficult to implement a timed wait_until_connected method correctly
> -----------------------------------------------------------------------------------
>
>                 Key: ZOOKEEPER-1446
>                 URL: https://issues.apache.org/jira/browse/ZOOKEEPER-1446
>             Project: ZooKeeper
>          Issue Type: Bug
>          Components: c client
>    Affects Versions: 3.4.3, 3.3.5
>            Reporter: Stephen Tyree
>            Priority: Minor
>         Attachments: ZOOKEEPER-1446.patch
>
>
> When using the C API, one might feel inclined to create a zookeeper_wait_until_connected method which waits for some amount for a connected state event to occur. The code might look like the following (didn't actually compile this):
> //------
> static pthread_mutex_t kConnectedMutex = PTHREAD_MUTEX_INITIALIZER;
> static pthread_cond_t kConnectedCondvar = PTHREAD_COND_INITIALIZER;
> int zookeeper_wait_until_connected(zhandle_t* zk, const struct timespec* timeout)
> {
>   struct timespec abstime;
>   clock_gettime(TIMER_ABSTIME, &abstime);
>   abstime->tv_sec += timeout->tv_sec;
>   abstime->tv_nsec += timeout->tv_nsec;
>   pthread_mutex_lock(&kConnectedMutex);
>   if (zoo_state(zk) == ZOO_CONNECTED_STATE) {
>     return 1;
>   }
>   pthread_cond_timedwait(&kConnectedCondvar, &kConnectedMutex, &abstime);
>   int state = zoo_state(zk);
>   return (state == ZOO_CONNECTED_STATE);
> }
> void zookeeper_session_callback(zhandle_t* zh, int type, int state, const char* path, void* arg)
> {
>   pthread_mutex_lock(&kConnectedMutex);
>   if (type == ZOO_SESSION_EVENT && state == ZOO_CONNECTED_STATE) {
>     pthread_cond_broadcast(&kConnectedCondvar);
>   }
> }
> //-----
> That would work fine (assuming I didn't screw anything up), except that pthread_cond_timedwait can spuriously wakeup, making you not actually wait the desired timeout. The solution to this is to loop until the condition is met, which might look like the following:
> //---
>   int state = zoo_state(zk);
>   int result = 0;
>   while ((state == ZOO_CONNECTING_STATE || state == ZOO_ASSOCIATING_STATE) && result != ETIMEDOUT) {
>     result = pthread_cond_timedwait(&kConnectedCondvar, &kConnectedMutex, &abstime);
>     state = zoo_state(zk);
>   }
> //---
> That would work fine, except the state might be valid and connecting, yet not ZOO_CONNECTING_STATE or ZOO_ASSOCIATING_STATE, it might be 0 or, as implemented recently courtesy of zookeeper-1108, 999. Checking for those states causes your code to rely upon an implementation detail of zookeeper, a problem highlighted by that implementation detail changing recently. Is there any way this behavior can become documented (via a ZOO_INITIALIZED_STATE or something like that), or is there any way this behavior can be supported by the library itself?

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira