Pass MSG_NOSIGNAL to sendmsg on OpenBSD

[nickg]

This avoids Asio applications exiting with unexpected SIGPIPEs on
OpenBSD. Note that FreeBSD has the SO_NOSIGPIPE socket option so doesn't
need this.

[namtsui]

I also came across this while getting a unit test (test_url_seed's url_seed_ssl) to pass on OpenBSD for libtorrent-rasterbar.

I had used a reduced version of this patch that doesn't cover all the sendmsg possibilities. This did allow the test to pass.

$OpenBSD$

Index: boost/asio/detail/impl/socket_ops.ipp
--- boost/asio/detail/impl/socket_ops.ipp.orig
+++ boost/asio/detail/impl/socket_ops.ipp
@@ -1178,7 +1178,7 @@ signed_size_type send(socket_type s, const buf* bufs,
   msghdr msg = msghdr();
   msg.msg_iov = const_cast<buf*>(bufs);
   msg.msg_iovlen = static_cast<int>(count);
-#if defined(__linux__)
+#if defined(__linux__) || defined(__OpenBSD__)
   flags |= MSG_NOSIGNAL;
 #endif // defined(__linux__)
   signed_size_type result = error_wrapper(::sendmsg(s, &msg, flags), ec);

@arvidn from libtorrent-rasterbar had suggested that this be submitted as a PR to boost asio. I have not done so. Thank you for submitting this.

explanation: arvidn/libtorrent#4211 (comment)
feedback from @arvidn: arvidn/libtorrent#4211 (comment)

[brad0]

This avoids Asio applications exiting with unexpected SIGPIPEs on
OpenBSD. Note that FreeBSD has the SO_NOSIGPIPE socket option so doesn't
need this.

I see NetBSD also has the socket option. Though the Asio code does not utilize it on
NetBSD at the moment.

[namtsui]

Based on these hints from the OpenBSD generic porting tips and Sascha Wildner from DragonflyBSD, #ifdef SO_NOSIGPIPE and #ifdef MSG_NOSIGNAL could be used instead to test directly for the features.

https://www.openbsd.org/faq/ports/guide.html#PortsGeneric

__OpenBSD__ should be used sparingly, if at all. Constructs that look like

#if defined(__NetBSD__) || defined(__FreeBSD__)

are often inappropriate. Don't add blindly __OpenBSD__ to it. Instead, try to figure out what's going on, and what actual feature is needed.

https://www.dragonflybsd.org/mailarchive/users/2011-03/msg00028.html

A saner way would be to replace those "#ifdef BSD" with "#ifdef SO_NOSIGPIPE" in order to check if the system allows turning off SIGPIPE for a socket. This is how for example BIND does it.

After @brad0's mention of NetBSD I checked all the BSDs. NetBSD and FreeBSD have SO_NOSIGPIPE. DragonflyBSD and OpenBSD do not have SO_NOSIGPIPE.

SO_NOSIGPIPE:
netbsd: https://man.netbsd.org/man.netbsd.org/setsockopt.2
freebsd: https://www.freebsd.org/cgi/man.cgi?query=setsockopt&apropos=0&sektion=2&manpath=FreeBSD+12.1-RELEASE+and+Ports&arch=default&format=html

dragonflybsd: https://man.dragonflybsd.org/?command=setsockopt&section=ANY
openbsd: https://man.openbsd.org/setsockopt.2

netbsd, freebsd, dragonflybsd and openbsd all have MSG_NOSIGNAL.

MSG_NOSIGNAL:
netbsd: https://man.netbsd.org/sendmsg.2
freebsd: https://www.freebsd.org/cgi/man.cgi?query=sendmsg&apropos=0&sektion=2&manpath=FreeBSD+12.1-RELEASE+and+Ports&arch=default&format=html
dragonflybsd: https://man.dragonflybsd.org/?command=sendmsg&section=ANY
openbsd: https://man.openbsd.org/sendmsg.2

[nickg]

MSG_NOSIGNAL was standardised in POSIX.1-2008 so I suppose the proper fix is to enable it based on _POSIX_VERSION. I've done this separately at nickg/asio@e19aa862022 and can update the pull request here if it's acceptable.

[pgroke-dt]

Fixing this for Solaris and AIX would also be highly welcome.

[chriskohlhoff]

Fixed in asio 1.18.1 / boost 1.75.