[ https://issues.apache.org/jira/browse/ZOOKEEPER-1519?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13587520#comment-13587520 ]
Daniel Lescohier commented on ZOOKEEPER-1519: --------------------------------------------- You're correct. I didn't have time to test the patch, but I wanted to get it out there for discussion. I looked further, and the callers get void* data from the public API with no length parameter. So, the public API does not allow us to copy the data. In order to fix it, it looks like a public API change is required. Either: 1. Document in the API that the caller cannot free that memory until the zookeeper library is done with it (which also means it can't be a pointer to memory on the stack). I assume that the library is done with it once it calls the completion callback? So the program can free it once it gets the same pointer back in a callback (or when the zookeeper connection is closed). I think this would make it hard to integrate with scripting languages like Python, because the scripting language C interface would have to copy the memory, account for it in some global structure, and free it once it sees that pointer again in a callback or when the zookeeper connection is closed. 2. Document in the API that the void * must be malloc'ed memory, and the ownership is passed to the library (which means the caller copies it, and the library frees it). That's also a difficult API. 3. Add a data length parameter to the API, so the library can copy it. 4. Don't use a void * for the 'data' parameter, use something else. > Zookeeper Async calls can reference free()'d memory > --------------------------------------------------- > > Key: ZOOKEEPER-1519 > URL: https://issues.apache.org/jira/browse/ZOOKEEPER-1519 > Project: ZooKeeper > Issue Type: Bug > Components: c client > Affects Versions: 3.3.3, 3.3.6 > Environment: Ubuntu 11.10, Ubuntu packaged Zookeeper 3.3.3 with some > backported fixes. > Reporter: Mark Gius > Attachments: zookeeper-1519.patch > > > zoo_acreate() and zoo_aset() take a char * argument for data and prepare a > call to zookeeper. This char * doesn't seem to be duplicated at any point, > making it possible that the caller of the asynchronous function might > potentially free() the char * argument before the zookeeper library completes > its request. This is unlikely to present a real problem unless the freed > memory is re-used before zookeeper consumes it. I've been unable to > reproduce this issue using pure C as a result. > However, ZKPython is a whole different story. Consider this snippet: > ok = zookeeper.acreate(handle, path, json.dumps(value), > acl, flags, callback) > assert ok == zookeeper.OK > In this snippet, json.dumps() allocates a string which is passed into the > acreate(). When acreate() returns, the zookeeper request has been > constructed with a pointer to the string allocated by json.dumps(). Also > when acreate() returns, that string is now referenced by 0 things (ZKPython > doesn't bump the refcount) and the string is eligible for garbage collection > and re-use. The Zookeeper request now has a pointer to dangerous freed > memory. > I've been seeing odd behavior in our development environments for some time > now, where it appeared as though two separate JSON payloads had been joined > together. Python has been allocating a new JSON string in the middle of the > old string that an incomplete zookeeper async call had not yet processed. > I am not sure if this is a behavior that should be documented, or if the C > binding implementation needs to be updated to create copies of the data > payload provided for aset and acreate. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira