[RIVER-254] request initiation can block on I/O for inapplicable connection

• Type: Bug
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: jtsk_2.0
• Fix Version/s: River_2.1.2
• Component/s: net_jini_jeri
• Bugtraq ID: 6307813
• Assignee: Peter Jones
• Reporter: John McClain
• Created: 17/Aug/07 17:05
• Updated: 06/Jan/10 06:04
• Resolved: 24/Apr/08 04:37

Description

Bugtraq ID 6307813

The following blocked state of several threads attempting to initiate a request to the same net.jini.jeri.connection-based endpoint has been observed (see attached partial thread dump):

(A) thread blocked in com.sun.jini.jeri.internal.mux.Mux.start waiting synchronously for the receipt of the ServerConnectionHeader message over a new connection while holding the lock for the ConnectionManager.OutboundMux instance for that connection

(B) thread holding the lock for the ConnectionManager for the endpoint while trying to determine the active and idle connections to pass to ConnectionEndpoint.connect, blocked waiting to acquiring the lock for the OutboundMux instance (in order to determine whether it is active or idle)

(C) many threads blocked waiting to acquire the lock for the ConnectionManager for the endpoint in order to initiate a new request to that endpoint

(D) the reaper thread for the ConnectionManager waiting to acquire the lock for the ConnectionManager in order to check for idle connections

For TcpEndpoint, this situation might not seem so bad, because effectively all of thread types (A), (B), and (C) are attempting to initiate a request to the same endpoint, and (A) has an established connection that should be usable for all of those threads (established connections to the same endpoint are effectively equal as far as TcpEndpoint is concerned-- constraints do not make one connection more or less suitable for a given request), so they all might as well wait for (A) to complete, successfully or otherwise-- there is not much reason to believe that the other threads would have better fortune were they to attempt to establish new connections concurrently with (A).

For other net.jini.jeri.connect-based endpoints, however, like SslEndpoint and KerberosEndpoint, constraints can make a given connection not applicable or not preferable for a given request, and therefore it is more clearly a bug that initiating a request can block on the receipt of a ServerConnectionHeader message for any other connection to the endpoint.