Move JGit off JSch?

Move JGit off JSch?

[Shawn O. Pearce]

I'm fed up with JSch.

The remote.<name>.timeout stuff I added recently to JGit causes
JSch to throw NPE on timeout[1] and the original author seems to
think that's a reasonable thing to do[2].

MINA SSHD[3] contains a different client implementation, licensed
under the Apache License 2.0.  Its currently missing public key
support[4] but the library is actually a pretty solid library,
and the author seems to have a decent clue.

I'm really starting to consider adding the public key support the
client side of MINA SSHD, and then rewriting TransportGitSsh to use
it instead of JSch.  The downside is, that really increases our
dependency chain, and SSH key management under Eclipse becomes a
unique EGit problem as we can't piggy-back onto the existing JSch
key management already done by the workbench.

But, I just spent the last ~3 hours digging around JSch trying to
fix this NPE bug... and realized the code is very thread unsafe,
uses at least 2 threads per connection (your thread, and a thread
it spawns for you) so you have no choice but to use threads, and
the author seems to think random NPEs are OK (see [2] and all of
the NPE bugs on sourceforge).

[1] http://jira.source.android.com/jira/browse/GERRIT-233
[2] http://www.mail-archive.com/jsch-users@lists.sourceforge.net/msg00520.html
[3] http://mina.apache.org/sshd/
[4] https://issues.apache.org/jira/browse/SSHD-25

-- 
Shawn.

Re: Move JGit off JSch?

[Ferry Huberts]

Shawn O. Pearce wrote:
> I'm fed up with JSch.

Good to know!

Last year I moved our application away from j2ssh to jsch because j2ssh
was pratically dead and rather oddly implemented within our application.
I chose jsch because Eclipse was using it and development was actually
being done on it. I wasn't aware of the issues with it though and also
was not aware of MINA.
I'll start looking into MINA soon because we're having some rather odd
issue in our application that seem to point to problems with jsch under
heavy load/many parallel connections. Your story lit up a light in my
head to that effect :-)

Are the Eclipse folks aware of these issues? they seem rather serious.
Wouldn't it be more effective to try to improve jsch through the eclipse
foundation/community, from multiple sides? (excuse me for asking the
obvious)

My background is in massive parallel systems so I know about the pain of
implementing this kind of thing correctly (and also the reward of seeing
it working...)

Re: Move JGit off JSch?

[Robin Rosenberg]

torsdag 25 juni 2009 20:14:28 skrev "Shawn O. Pearce" <spearce@spearce.org>:
> I'm fed up with JSch.
> 
> The remote.<name>.timeout stuff I added recently to JGit causes
> JSch to throw NPE on timeout[1] and the original author seems to
> think that's a reasonable thing to do[2].
> 
> MINA SSHD[3] contains a different client implementation, licensed
> under the Apache License 2.0.  Its currently missing public key
> support[4] but the library is actually a pretty solid library,
> and the author seems to have a decent clue.
Clues are useful :)

> I'm really starting to consider adding the public key support the
> client side of MINA SSHD, and then rewriting TransportGitSsh to use
> it instead of JSch.  The downside is, that really increases our
> dependency chain, and SSH key management under Eclipse becomes a
> unique EGit problem as we can't piggy-back onto the existing JSch
> key management already done by the workbench.

I haven't really looked into JSch internals, but rewriting complex code
without trying hard to fix the problem might not be worth it. New
code /will/ have problems and having a different mechanism for key
management is in some very obscure corner on the reverse side of
my wish list. 

> But, I just spent the last ~3 hours digging around JSch trying to

I'm not quite as efficient as you, but I've spent more than three
hours on fixing much simpler things.

> fix this NPE bug... and realized the code is very thread unsafe,
> uses at least 2 threads per connection (your thread, and a thread
> it spawns for you) so you have no choice but to use threads, and
> the author seems to think random NPEs are OK (see [2] and all of
> the NPE bugs on sourceforge).

Can't we bug him a little more, but then again patches are probably
more convincing than bug reports. Perhaps you could explain the 
issues in more detail before we start rewriting the key management
into SSHD. 

Should we raise a bug report in eclipse about replacing
JSch, in order to get some feedback from Eclipse people?

-- robin