ServerSocket.accept()/scheduler bug (Kaffe 0.91 BSDI 3.1)
dpcaldwell at sherwin.com
Thu Feb 5 09:46:03 PST 1998
______________________________ Forward Header __________________________________
Subject: ServerSocket.accept()/scheduler bug (Kaffe 0.91 BSDI 3.1)
Author: David Caldwell at TRANS_PO
Date: 2/5/98 8:48 AM
I tried to subscribe to the mailing list by sending a "subscribe"
message to kaffe-request at tjwassoc.demon.co.uk, but I haven't heard
back. This message is being sent to kaffe at tjwassoc.demon.co.uk.
After long examination, I believe I've found a Kaffe bug. The method
accept() of java.net.ServerSocket not only blocks when it is called,
(a) ignore the SoTimeout parameter,
(b) never relinquish the CPU.
My program runs fine on Sun's Windows 95 implementation, which has a
pre-emptive scheduler. Though I have not been able to test the
program with a non-preemptive scheduler, I've done enough testing to
decide that it -should- work on a non-pre-emptive scheduler, also.
I've tried a number of workarounds -- killing the thread, lowering its
priority, etc. You can't kill the thread because it's blocking and
won't give any other thread the CPU for even one instruction. No
matter how low the thread's priority is, if it gets to the accept()
call, it will hang there forever -- or until another socket connection
Because it will release when another connection is received, I could
construct a workaround -if- there was a ServerSocket method which told
whether there were any connections backlogged (analogous to the
ready() method for InputStreams). But I can't find any such method --
so that's a Java question.
ACTUAL (BUT UNTESTED, NATCH) WORKAROUND
The other workaround (which I haven't tried -- but am counting on) is
to start a SECOND virtual machine and have it (more or less)
continually request socket connections with the first VM; then send
something through the socket verifying that it's a "fake" connection,
thus allowing the first VM to unblock and proceed. Inelegant, but if
you're desperate, it would probably work. I'll follow up.
I can't actually look at the source code because I don't control my
own Kaffe distribution (my web hoster does). Unless I want to
download it locally, which I don't, at least for the moment.
Also, I'm not sure that I'm actually subscribed to the mailing list
(haven't received confirmation) so any replies should go to
inonit at inonit.com.
WHAT I'M LOOKING FOR
(a) Workaround, if possible.
(b) Less important would be a fix to the underlying problem.
More information about the kaffe