Andrew Kyle Purtell created ZOOKEEPER-5123:
----------------------------------------------
Summary: (C client) Broadcast completion of auth can return ZOK
for rejected credentials
Key: ZOOKEEPER-5123
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-5123
Project: ZooKeeper
Issue Type: Bug
Components: c client
Reporter: Andrew Kyle Purtell
Assignee: Andrew Kyle Purtell
When the C client receives a reply to a SetAuth request,
{{auth_completion_func}} runs every pending {{zoo_add_auth}} callback with the
status of that one reply. It never checks which credential the reply belongs
to. The client also puts SetAuth frames at the front of the send queue, so
frames from back-to-back calls go out in reverse order.
Suppose an application adds an invalid credential and then a valid one. The
server accepts the valid one first, and both callbacks receive {{{}ZOK{}}}. The
server then rejects the invalid credential and closes the connection. By that
time the application has already been told the rejected credential succeeded,
and it may act on that.
There are related problems on the same path:
* Credentials that the client sends again during a reconnect can complete
newly issued {{zoo_add_auth}} calls before the server has checked them.
* An auth reply on a handle with no credentials dereferences a NULL list and
crashes the I/O thread.
* Concurrent {{zoo_add_auth}} calls can both send the last credential in the
list, so one thread's credential is never sent.
* If the frame cannot be queued, {{zoo_add_auth}} returns an error but leaves
the callback registered.
The fix is to give each handle a FIFO of SetAuth frames waiting for a reply,
kept in the same order as the frames on the wire. The server processes SetAuth
packets inline and replies in the order it receives them. So each reply
completes only the oldest pending entry and updates only that entry's
credential. A rejection still moves the handle to
{{{}ZOO_AUTH_FAILED_STATE{}}}, and the teardown fails any callbacks still
pending.
The fix also address the related problems:
* An auth reply with no pending frame is logged and ignored.
* {{zoo_add_auth}} holds the auth lock while it queues its own credential's
frame, and it checks its memory allocations.
* {{zoo_add_auth}} now adds its frame to the end of the send queue, as the
Java client does.
* {{zoo_add_auth}} no longer sends a frame while the session is still being
established; the credential is sent once the session is ready. Before, that
path sent the credential twice.
A new {{Zookeeper_authCompletion}} test suite covers these cases.
The public API and ABI are unchanged.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)