is setsockopt SO_NOSIGPIPE work?

[Bruce M. Simpson]

Anton Yuzhaninov wrote:
> RE> It works, but only if you use send() instead of write().
> RE> Alternatively, you can control the behavior on a per
> RE> message basis, by passing the MSG_NOSIGNAL in the "flags"
> RE> argument to the send() call (without having to set a
> RE> socket option).
>
> Thanks, with send() it works fine.
> I think it should be documented in setsockopt(2).
>   
AFAIK this is not a POSIX sockopt. I can only trace it back to MacOS X 
as the origin.
Most applications I know of set the handler for SIGPIPE to SIG_IGN in 
such situations.

Call graph: write() -> dofilewrite() -> soo_write() -> pru_send()

Looking at the code for the generic write() path it looks like we would 
never squelch this kind of SIGPIPE intentionally.
In soo_write() we check the SO_NOSIGPIPE option to tell if we should 
call psignal().
However, as soon as we return from soo_write(), the EPIPE is mapped to 
psignal() by the generic code in dofilewrite() which generates the 
SIGPIPE you are seeing.

I think this may be a bug but in the absence of precise written 
requirements I can't be sure. :-)

BMS