I can confirm that poll() on Mac OS X 10.4 is broken in some ways. The
issue affecting curl is that poll() sometimes gives POLLHUP where
other implementations give POLLIN.

A bug has been filed in Apple's bug tracking system, and is being
actively investigated. Email me privately if you have any other
information to report regarding this issue. This address or 'toby at
apple dot com' are both fine.

Toby Peterson
Software Engineer, Apple Computer

On 5/13/05, Daniel Stenberg <daniel-curl_at_haxx.se> wrote:
> On Fri, 13 May 2005, Grigory Entin wrote:
>
> > Well, just in case, I'm not sure that that's bad "poll" that causes the
> > malfunction - it just happens that switching back to old schema solves the
> > problem.. But you know it better. ;)
>
> The poll() way works on numerous other platforms, which I think is proof
> enough to say that Mac OS X 10.4 has a bad poll().
>
> > And, I guess, you know that that configure test under 10.3 fails because of
> > the *assertion* in system poll on NULL argument (I'm not sure if the
> > assertion is the expected result).
>
> The configure test for poll() is expected to return 0 since poll(0,0,10)
> should return a timeout after 10ms.
>
> > I just did it (with ktrace): yes, the built-in version doesn't use poll, but
> > uses select.
>
> Thanks. I've now comitted a modified configure test that skips the poll() test
> if running on darwin. Would you be able to test it? It is available in CVS
> right now, and in the soon (in about 4 hours and 30 minutes) appearing daily
> snapshot for May 14th: http://cool.haxx.se/curl-daily/
>
> --
> Commercial curl and libcurl Technical Support: http://haxx.se/curl.html
>
Received on 2005-05-13