On 6/2/09, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Jeremy Kerr <jk@ozlabs.org> writes:
>
> >> The consideration is that the application fails completely on server
> >> disconnect (because it gets SIGPIPE'd). This was long ago deemed
> >> unacceptable, and we aren't likely to change our opinion on that.
>
> > OK, understood. I'm guessing MSG_NOSIGNAL on the send() isn't portable
> > enough here?
>
>
> Well, it's certainly not 100% portable, but I wouldn't object to a patch
> that tests for it and uses it where it works.
>
> One question that might be a bit hard to answer is whether mere
> existence of the #define is sufficient evidence that the feature works.
> We've had problems before with userland headers not being in sync
> with what the kernel knows.
Well, we could just test in configure perhaps? Runtime test is also
possible (if kernel gives error on unknown flag). Safest would
be enable on known-good OSes, maybe with version check?
--
marko