Re: Retrying close() after error on Linux !?

[Bruno Haible]

Hi Eric,

> >> I think we'll need to fall back to a #if based on platforms.
> > 
> > If I'm reading the POSIX bug correctly, HPUX is the only system where 
> > fds are left open on EINTR.

Early Solaris is definitely in this camp as well.

> > Seems like a reasonable enough #if when 
> > coupled with good comments pointing back to the POSIX discussion.
> 
> One other possible alternative: instead of calling close(fd) on a random 
> unknown fd, and worrying about the non-portability of EINTR behavior 
> when that fd is something that takes a long time to close, you could use 
> dup2()'s semantics to replace the original fd with something known to be 
> safe to close (for example, closing an fd open on /dev/null doesn't have 
> the risk of EINTR like you would have with closing an fd open on a pty 
> or tape device).

This is an interesting, creative idea. However, because few people use this
'dup2 as a kind of atomic close' workaround, I fear that it might lead us
into the proximity of OS bugs. Hardly anyone has tested how dup2 works
in the presence of signals and ttys, in the past.

Bruno