Jun-ichiro itojun Hagino <itojun@itojun.org>:

Now first check if IPv6 is supported, then use PF_UNSPEC. If not, use PF_INET.
It'll solve both the "slow name lookup" problem on IPv4 and still work fine on
IPv6 hosts.

Bug report #530204 has more details:
http://sourceforge.net/tracker/?func=detail&atid=100976&aid=530204&group_id=976

diff --git a/lib/hostip.c b/lib/hostip.c
index 412844c3bbedc80b6c70e04c26258719761eb61c..fe18e1d40e5aab5323fcb8c53199f91d58b3af0f 100644 (file)
--- a/lib/hostip.c
+++ b/lib/hostip.c
@@ -291,9 +291,23 @@ Curl_addrinfo *Curl_getaddrinfo(struct SessionHandle *data,
   struct addrinfo hints, *res;
   int error;
   char sbuf[NI_MAXSERV];
-
+  int s, pf = PF_UNSPEC;
+
+  /* see if we have an IPv6 stack */
+  s = socket(PF_INET6, SOCK_DGRAM, 0);
+  if (s < 0)
+    /* Some non-IPv6 stacks have been found to make very slow name resolves
+     * when PF_UNSPEC is used, so thus we switch to a mere PF_INET lookup if
+     * the stack seems to be a non-ipv6 one. */
+    pf = PF_INET;
+  else
+    /* This seems to be an IPv6-capable stack, use PF_UNSPEC for the widest
+     * possible checks. And close the socket again.
+     */
+    close(s);
+ 
   memset(&hints, 0, sizeof(hints));
-  hints.ai_family = PF_UNSPEC;
+  hints.ai_family = pf;
   hints.ai_socktype = SOCK_STREAM;
   hints.ai_flags = AI_CANONNAME;
   snprintf(sbuf, sizeof(sbuf), "%d", port);