Re: Runtime.exec() implementation

[Archie Cobbs]

Andrew Haley wrote:
>  > I wrote an implementation of Runtime.exec() (including a new class
>  > java.lang.VMProcess which extends java.lang.Process and native code)
>  > for JC. This can be easily merged into Classpath if people want it.
>  > It seems like a pretty glaring omission.
>  > 
>  > This uses the usual Unix functions of fork(), exec(), waitpid(), etc.
>  > and should be as portable as the existing native code.  I'd have to
>  > rewrite my JCNI native code in JNI but that would be easy.
>  > 
>  > Let me know if people are interested.
> 
> This sounds good.  However, the precise implementation of this class
> is very delicate and error-prone, and we found a number of bugs in
> gcj's implementation.  The trouble is that our bugs were only noticed
> after some time.  I guess you may write fewer bugs than we did, but
> there still might be one or two in yours.  :-)

I'm certain there are bugs in mine :-) I just implemented things in
the "obvious" way.

> Would you be interested in having a look at the libgcj implementation
> and comparing?  Rather than having to download or check out all of
> libgcj, I could just send you the few files.

Yes, please do send me the files as I'd be interested to see them.

One thing I noticed is that waitFor() can throw InterruptedException,
which to me implies that they are implementing it via Object.wait()
or Thread.join().

For example, creating a new thread to fork() and then waitpid() on
each process would be heavyweight but might work well. Then you just
join() the thread in waitFor(), plus problem #2 goes away.

-Archie

__________________________________________________________________________
Archie Cobbs      *        CTO, Awarix        *      http://www.awarix.com