Ken Giusti created DISPATCH-2249:
------------------------------------
Summary: Refactor router shutdown to eliminate leaks and races
Key: DISPATCH-2249
URL: https://issues.apache.org/jira/browse/DISPATCH-2249
Project: Qpid Dispatch
Issue Type: Task
Components: Router Node
Affects Versions: 1.17.0
Reporter: Ken Giusti
Assignee: Michael Goulish
Fix For: Backlog
This work item refactors the router shutdown sequence.
On shutdown the router currently simply terminates all threads (but the main
thread) mid-process. All state is essentially frozen at the point where the
shutdown signal is received. Then the main thread attempts to clean up state
before exiting.
This approach is error prone and results in memory leaks (see open JIRAs). In
addition it requires a bespoke cleanup handler that essentially duplicates
run-time cleanup code (e.g. link close handling, connection close handling,
etc).
It would be better to implement a controlled shutdown that leverages the
"normal" connection close/management delete code that exists in the router.
For example, the new shutdown process could go something like this:
Add two new attributes to the "router" management entity: adminStatus and
operStatus:
adminStatus values: ["up", "down"]
operStatus: ["active", "quiescing", "shutdown"]
adminStatus defaults to "up". When modified either via management or
SIGTERM/QUIT/INT/etc) to "down" initiate the shutdown process:
# Set operStatus to quiescing
# Close & delete all listeners and connectors. This will prevent new
connections from being established.
# Initiate close of all active connections
# Wait for all connections to complete close and delete
# Join() all I/O threads but the main thread (this leaves the main thread and
the core thread).
# Issue a new action to the core thread to cause it to clean up its resources
and exit
# Join() the core thread.
# Clean up any remaining server state then exit the main thread.
This is just an example of a possible shutdown sequence. A proper design
document should be proposed for review as a first step.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
