Fixes for OpenBSD

Fixes for OpenBSD

[Jeremy Evans]

I'm not sure if these issues are specific to OpenBSD or not, but they
are required on OpenBSD.

With these two patches, kgio-2.3.3 passes the included tests on OpenBSD
except for one failure:

1) Failure:
test_wait_writable_set(TestKgioTcpConnect) [test/test_tcp_connect.rb:69]:
<"waited"> expected but was
<nil>.

Hope this helps,
Jeremy

Fix kgio_peek by adding an empty kgio_autopush_recv
if !KGIO_NOPUSH.

--- ext/kgio/autopush.c.orig	Thu May  5 11:02:08 2011
+++ ext/kgio/autopush.c	Thu May  5 11:02:12 2011
@@ -242,6 +242,7 @@ static void push_pending_data(VALUE io)
 		rb_sys_fail("setsockopt(TCP_CORK/TCP_NOPUSH, 1)");
 }
 #else /* !KGIO_NOPUSH */
+void kgio_autopush_recv(VALUE io){}
 void init_kgio_autopush(void)
 {
 }

getaddrinfo gives an error if you use AF_UNSPEC as the ai_family,
the man page says PF_UNSPEC should be used.  And other entries
in hints need to be 0 or NULL.  

--- ext/kgio/connect.c.orig	Wed Dec 31 16:00:00 1969
+++ ext/kgio/connect.c	Thu May  5 10:47:10 2011
@@ -69,7 +69,8 @@ static VALUE tcp_connect(VALUE klass, VALUE ip, VALUE 
 	rc = snprintf(ipport, sizeof(ipport), "%u", uport);
 	if (rc >= (int)sizeof(ipport) || rc <= 0)
 		rb_raise(rb_eArgError, "invalid TCP port: %u", uport);
-	hints.ai_family = AF_UNSPEC;
+	memset(&hints, 0, sizeof(hints));
+	hints.ai_family = PF_UNSPEC;
 	hints.ai_socktype = SOCK_STREAM;
 	hints.ai_protocol = IPPROTO_TCP;
 	/* disallow non-deterministic DNS lookups */

Re: Fixes for OpenBSD

[Eric Wong]

Jeremy Evans <code@jeremyevans.net> wrote:
<snip> other changes look fine
> getaddrinfo gives an error if you use AF_UNSPEC as the ai_family,
> the man page says PF_UNSPEC should be used.  And other entries
> in hints need to be 0 or NULL.

I'm not sure why OpenBSD differs from POSIX here:
  http://pubs.opengroup.org/onlinepubs/009695399/functions/getaddrinfo.html

Even though PF_UNSPEC == AF_UNSPEC in most systems, I'd rather go with
the following patch and single out OpenBSD (and maybe other *BSDs) since
"AF_" (address family) makes more sense than "PF_" (protocol family):

diff --git a/ext/kgio/connect.c b/ext/kgio/connect.c
index 3b40054..e4ee9fb 100644
--- a/ext/kgio/connect.c
+++ b/ext/kgio/connect.c
@@ -1,6 +1,21 @@
 #include "kgio.h"
 #include "sock_for_fd.h"
 
+/*
+ * OpenBSD getaddrinfo() requires PF_UNSPEC, but POSIX (and Linux) docs
+ * say AF_UNSPEC.  AF_UNSPEC is typically PF_UNSPEC anyways, but apparently
+ * OpenBSD differs.
+ *
+ * http://www.openbsd.org/cgi-bin/man.cgi?query=getaddrinfo&apropos=0&sektion=0&manpath=OpenBSD+Current&arch=i386&format=html
+ * http://pubs.opengroup.org/onlinepubs/009695399/functions/getaddrinfo.html
+ */
+#if defined(__OpenBSD__)
+#  if defined(AF_UNSPEC)
+#    undef AF_UNSPEC
+#  endif
+#  define AF_UNSPEC PF_UNSPEC
+#endif
+
 static void close_fail(int fd, const char *msg)
 {
 	int saved_errno = errno;
@@ -69,6 +84,7 @@ static VALUE tcp_connect(VALUE klass, VALUE ip, VALUE port, int io_wait)
 	rc = snprintf(ipport, sizeof(ipport), "%u", uport);
 	if (rc >= (int)sizeof(ipport) || rc <= 0)
 		rb_raise(rb_eArgError, "invalid TCP port: %u", uport);
+	memset(&hints, 0, sizeof(hints));
 	hints.ai_family = AF_UNSPEC;
 	hints.ai_socktype = SOCK_STREAM;
 	hints.ai_protocol = IPPROTO_TCP;

-- 
Eric Wong

Re: Fixes for OpenBSD

[Jeremy Evans]

> I'm not sure why OpenBSD differs from POSIX here:
>   http://pubs.opengroup.org/onlinepubs/009695399/functions/getaddrinfo.html
> 
> Even though PF_UNSPEC == AF_UNSPEC in most systems, I'd rather go with
> the following patch and single out OpenBSD (and maybe other *BSDs) since
> "AF_" (address family) makes more sense than "PF_" (protocol family):
 
It looks like the AF_UNSPEC -> PF_UNSPEC is not necessary, they must be
defined to the same thing.  Why the OpenBSD man page recommends
PF_UNSPEC, I'm not sure.  It looks like the memset was the necessary
fix.

Thanks,
Jeremy

Re: Fixes for OpenBSD

[Eric Wong]

Jeremy Evans <code@jeremyevans.net> wrote:
> > I'm not sure why OpenBSD differs from POSIX here:
> >   http://pubs.opengroup.org/onlinepubs/009695399/functions/getaddrinfo.html
> > 
> > Even though PF_UNSPEC == AF_UNSPEC in most systems, I'd rather go with
> > the following patch and single out OpenBSD (and maybe other *BSDs) since
> > "AF_" (address family) makes more sense than "PF_" (protocol family):
>  
> It looks like the AF_UNSPEC -> PF_UNSPEC is not necessary, they must be
> defined to the same thing.  Why the OpenBSD man page recommends
> PF_UNSPEC, I'm not sure.  It looks like the memset was the necessary
> fix.

OK, thanks!

-- 
Eric Wong

Re: Fixes for OpenBSD

[Eric Wong]

Jeremy Evans <code@jeremyevans.net> wrote:
> 1) Failure:
> test_wait_writable_set(TestKgioTcpConnect) [test/test_tcp_connect.rb:69]:
> <"waited"> expected but was
> <nil>.

Ah, some TCP stacks may optimize loopback connections so non-blocking
connect() won't return with EAGAIN (like with Unix sockets).  I'll just
disable the test on non-Linux.

> Hope this helps,

Thank you very much for your feedback :)

-- 
Eric Wong