Tudor Bosman created ZOOKEEPER-3868:
---------------------------------------
Summary: Completion leak in zookeeper_close
Key: ZOOKEEPER-3868
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3868
Project: ZooKeeper
Issue Type: Bug
Components: c client
Affects Versions: 3.5.8, 3.6.1, 3.7.0
Reporter: Tudor Bosman
Using the multithreaded C library.
Due to a race condition between `zookeeper_close` and `do_completion`, it is
possible for `zookeeper_close` to close the handle without running the pending
completions, causing a completion leak (not calling the client's completion
callbacks) and a memory leak.
`zookeeper_close` sets `close_requested` here:
[https://github.com/apache/zookeeper/blob/master/zookeeper-client/zookeeper-client-c/src/zookeeper.c#L3752];
after that point, the completion thread can exit at any time:
[https://github.com/apache/zookeeper/blob/8da9c723ac1a889c989ecefada722ed858049537/zookeeper-client/zookeeper-client-c/src/mt_adaptor.c#L473]
But, even after `zookeeper_close`, completions can still be added to
`completions_to_process` from two different places: the IO thread (which is
still running) and `zookeeper_close` itself, via `free_completions`:
[https://github.com/apache/zookeeper/blob/master/zookeeper-client/zookeeper-client-c/src/zookeeper.c#L3760]
I have a fix (I'll update this issue with the pull request) that uses a
separate `terminate_completion` flag in `adaptor_threads` instead of
`zh->close_requested` to make the IO thread exit.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
