[jboss-jira] [JBoss JIRA] (WFCORE-1616) ServerSuspendHandler tries to do step completion and rollback handling in a different thread

[Brian Stansberry (JIRA)]

    [ https://issues.jboss.org/browse/WFCORE-1616?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13256031#comment-13256031 ] 

Brian Stansberry commented on WFCORE-1616:
------------------------------------------

[~swd847] I'm not sure what semantics you wanted here. Was it to have the management op block until suspend is done (however long that may be), and thus support rolling back regardless of how long suspend takes? Or was it to support rollback only for the 1) case, where the suspend is immediate?

Blocking sounds better, although potentially indefinite blocking with the management lock held is a concern.

> ServerSuspendHandler tries to do step completion and rollback handling in a different thread
> --------------------------------------------------------------------------------------------
>
>                 Key: WFCORE-1616
>                 URL: https://issues.jboss.org/browse/WFCORE-1616
>             Project: WildFly Core
>          Issue Type: Bug
>          Components: Domain Management
>            Reporter: Brian Stansberry
>
> ServerSuspendHandler is calling OperationContext.completeStep() and passing in a rollback handler from a org.jboss.as.server.suspend.OperationListener that may be invoked by a different thread. The OperationContext is not intended to be invoked from multiple threads in this way. 
> 3 things can happen with this setup:
> 1) There is no activity preventing suspend, so the suspend controller immediately calls back to the OperationListener on the thread that's handling the operation. So then things work fine.
> 2) There is something that prevents synchronous suspend (i.e. user activity) so then the OperationListener gets invoked later by another thread. The completeStep call registers a rollback handler that will never get called because the op is already done. No harm, no foul unless the operation rolled back.
> 3) Same as 2) but the OperationListener gets invoked later by another thread while the mgmt op is still executing and somehow something goes wrong, since AbstractOperationContext.Step is not written for concurrent modification.



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)