On Mon, 29 Jan 2007, Daniel Johnson wrote:

> I've noticed a couple of things during my builds. One is that tests
> 518 and 537 always fail, and apparently always have failed on OS X.
> It's probably due to the need to use select instead of poll. Perhaps
> those tests should be disabled on OS X?

Your own submitted autobuild log shows: "checking if we are Mac OS X
(to disable poll)... yes". So it is clear that libcurl & curl have
been built using the working select() instead of the poll() function.
If built with a non reliable poll(), then other test cases would also
be failing besides 518 and 537.

I don't agree with disabling tests 518 and 537 on OS X.

In this specific case Its failure is indicating that in case that more
than FD_SETSIZE file descriptors, or sockets, are open libcurl or any
other program using select() won't be able to use a new connection.

The difference between libcurl and many other third party programs is
that when the above condition is met libcurl will behave properly and
won't actually perform the call to select(), since this could trigger
quite nasty effects.

Gory details about FD_SETSIZE and select() at
http://developer.apple.com/documentation/Darwin/Reference/Manpages/man5/types.5.html
and http://developer.apple.com/documentation/Darwin/Reference/Manpages/man2/select.2.html

Maybe these tests could be 'enhanced' when using select() instead of
poll(). Maybe even tests in configure script for a working select()
with more than FD_SETSIZE fds...

Anyway. Could you test if Mac OS X 10.4.8 implementation of poll() is
fine ? Edit configure.ac, replace single appearance of
disable_poll="yes"; with disable_poll="no"; rerun buildconf, configure
, make, and tests

Maybe current implementation of poll() is OK.

-- 
-=[Yang]=-