Bugzilla – Full Text Bug Listing
| Summary: | jgit hangs waiting for ssh error stream to be closed | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Product: | [Technology] JGit | Reporter: | Dmitry Neverov <dmitry.neverov> | ||||||
| Component: | JGit | Assignee: | Project Inbox <jgit.core-inbox> | ||||||
| Status: | RESOLVED FIXED | QA Contact: | |||||||
| Severity: | normal | ||||||||
| Priority: | P3 | CC: | sop | ||||||
| Version: | 0.7.1 | ||||||||
| Target Milestone: | 0.8.0 | ||||||||
| Hardware: | PC | ||||||||
| OS: | Linux | ||||||||
| Whiteboard: | |||||||||
| Attachments: |
|
||||||||
|
Description
Dmitry Neverov
Created attachment 168509 [details]
possible fix
(In reply to comment #0) > In close() method of SshFetchConnection and SshPushConnection > errorThread.join() can wait forever if jsch will not close channel's error > stream. Maybe join with timeout should be used. Path with possible fix is > attached Ick. IIRC we have to wait for the errorThread because closing stdout causes JSch to immediately tear down the command channel, which means errors can get lost if they weren't yet consumed by the error thread. In a case like this where the client side has aborted, there is no other way for JSch to know that we aren't going to continue. What about an alternative patch that tells the errorThread we are done and wish to terminate now? We might be able to use a smaller timeout on the join, so we have less time to wait for the close to finish. But the smaller timeout may cause us to miss an error message, so I'm wondering if we shouldn't also keep track of the connection state. Once the remote side has started to send us data its less likely to fail, so any close attempt after that means we probably won't need to wait for stderr data. http://egit.eclipse.org/r/708 has a counter-proposal patch based on your idea. Created attachment 168700 [details]
simplified halt patch
Taking into account bug https://bugs.eclipse.org/bugs/show_bug.cgi?id=313082, your patch in StreamCopyThread class can be simplified to attached one. Yes we can miss some error messages, but tracking state of connection will not help us, because jsch will never close error thread in case of exception in jgit code. I think if we made decision to close connection we have already read everything we want or we got exception and can throw it up to client code, so it is ok to miss some error messages. Anyway longer timeout won't solve this, because we can get errors after any timeout we choose. Committed as 2e961989e42b1fe7e8bd9eaa7a3d2e88a0d1d001 |