Shibboleth Developers

[Scott Cantor <>]

Ok, to bring folks up to speed on this problem Derek and I have been looking
into, I found a few days ago that the Shib target had stopped working on
sunra at some point recently. Specifically, calls to select() on the shar's
listener socket were failing.

The sunra box has one key difference from at least one box I know the target
still works on, here at OSU. It's got an entry in /etc/system that overrides
the default FD_SETSIZE and increases it from 1024 to 4096. If you look, the
header macro for FD_SETSIZE is still 1024, but getdtablesize() will return
4096, which is what the kernel reports, as it should.

I think that the larger size is blowing up select(), and I have some
evidence for that in some comments in some man pages that seem to be saying
that select() isn't designed to handle more than 1024, and has been
deprecated in favor of poll(). I see similar comments in the Solaris RPC
manpages.

Of course, we never could get Sun's ONC library to work, so we're using the
MIT port. And it uses getdtablesize() in some spots, so it's using the 4096
number too, and if I force my way past the select() call by passing 1024 as
the fdsize directly, the ONC calls crash anyway.

So, I tried commenting out the enlarged setting in /etc/system and rebooting
sunra. Guess what? It works now.

So....me not Unix guy, but obviously this is an "issue". With our code, I'd
have to say, since this should be up to the sysadmin to decide on the value.
But I defer to Derek or others who don't #include <winsock.h> for a living
like me...

-- Scott

------------------------------------------------------mace-shib-design-+
For list utilities, archives, subscribe, unsubscribe, etc. please visit the
ListProc web interface at 

    http://archives.internet2.edu/

------------------------------------------------------mace-shib-design--