Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
3.0.0, 3.4.13
-
Patch
Description
As reported (but never fixed) in the past by ZOOKEEPER-2409, ZOOKEEPER-2038 and (partly) ZOOKEEPER-2878, the C client lock-recipe implementation is broken.
I identified three issues.
The main one (as also reported in the aforementioned reports) is that the logic that goes through the lock waiting list is broken. child_floor uses strcmp and compares the full node name (i.e., sessionID-sequence) rather than only comparing the sequence number. This makes it possible for two different clients to hold the lock at the same time: assume two clients, one associated with session A, the other with session B, with A < B lexicographically. Now assume that at some point a thread in B holds a lock and a thread in A tries to acquire the same lock. A will manage to get the lock because of the wrong comparison function, so now two guys hold the lock.
The second issue is a possible deadlock inside zkr_lock_operation. zkr_lock_operation is always called by holding the mutex associated to the client lock. In some cases, zkr_lock_operaton may decide to give-up locking and call zkr_lock_unlock to release the lock. When this happens, it will try to acquire again the same phtread mutex, which will lead to a deadlock.
The third issue relates to the return value of zkr_lock_lock. According to the API docs, the functions returns 0 when no errors. Then it is up to the invoker to check when the lock is held by calling zkr_lock_isowner. However, the implementation, in case of no error, returns zkr_lock_isowner. This is wrong because it becomes impossible to distinguish an error condition from a success (but not ownerhsip). Instead the API (as described in the docs, btw) should return always 0 when no errors occur.
Shortly I will add the link to a PR fixing the issues.
