Andrew Kyle Purtell created ZOOKEEPER-5126:
----------------------------------------------
Summary: (C client) Async watch removals corrupt context and
report false success
Key: ZOOKEEPER-5126
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-5126
Project: ZooKeeper
Issue Type: Bug
Components: c client
Reporter: Andrew Kyle Purtell
Assignee: Andrew Kyle Purtell
The C client's watch removal APIs have three related defects.
First, in the multithreaded library, an asynchronous local removal through
{{zoo_aremove_watches}} or {{zoo_aremove_all_watches}} treats the caller's
context pointer as an internal synchronous-completion record. It locks a mutex,
sets flags and signals a condition variable inside application memory. This
corrupts the application's data and can hang the caller in
{{{}pthread_mutex_lock{}}}.
Second, the public header declares the completion parameter of both
asynchronous functions as a pointer to a function pointer, not as a function
pointer. A correctly typed callback is rejected by C++ compilers such as GCC 14
and later.
Third, when the caller passes a NULL watcher (the remove-all case), the client
treats the request as a match before it checks whether the path has any
watches. A local remove-all then returnsĀ {{ZOK}} for a path with no watches,
when the header promises {{ZNOWATCHER}} and the Java client throws
{{{}NoWatcherException{}}}.
All three defects are in the same API entry points, and they should be fixed
together.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)