Re: Bug#676461: libncursesw5: buggy test causes select() to be used instead of poll(), making ncurses really slow

[shawn]

resend, seems like i got blocked from the other address, even @ BTS

On Thu, 2012-06-07 at 06:51 -0400, Thomas Dickey wrote: 
> On Thu, Jun 07, 2012 at 06:42:24AM +0000, shawn wrote:
> > Package: libncursesw5
> > Version: 5.9-8
> > Severity: important
> > Tags: upstream
> > 
> > int main() {
> >         struct pollfd myfds;
> >         int ret;
> > 
> >         myfds.fd = 0;
> >         myfds.events = POLLIN;
> > 
> >         ret = poll(&myfds, 1, 100);
> 
> Reflecting on it, the check relies on the given file-descriptor and timeout.
> Standard input is not redirected.
> 
> The check does need a tty though (for Darwin).  The return-code for that might
> be a -1 (making a comparison for that case more apt - will have to test).
> 
> Almost all of my build-logs (reviewing my build-logs, I see one out of ~60
> combinations failing for Linux) for Linux show it working - probably depends
> upon the environment.
All of my computers' ncurses, under a number of distributions, are using
select(). 
> 
> > If i compile the C test myself, adding a printf at the end, ret==0 after 
> > the poll().
> > 
> > The buildds are returning 
> > 
> > "checking if poll really works... no"
> > 
> > yet Linux's poll() is fine.
> 
> usually (I seem to recall that the original reason for testing poll vs
> select was in fact a buggy implementation on Linux - that's been a while).

I recompiled ncurses forcing poll() and it worked fine with my existing
ncurses binaries; on 3.2 kernel. Copying these libncurses and libtinfo,
plus new libc, to a Centos 6.2 machine, (2.6.32) and using it with
system binaries there also worked, with is sadly the oldest OS I have
with disk space available.

-- 
-Shawn Landden