Lighttpd

Ok, imagine we had an option "listen-on-ipv6-too":

server.bind = "127.0.0.1" 
listen-on-ipv6-too = "enable" 

Do we listen to "[::]" now or "[::1]" ?

I don't like any magic in that area (yes, i don't like $SERVER["socket"] either).

I don't see a good way to provide this option...

Well, there are three questions when binding: IP address, port and transport (IPv4, IPv6).
IMO the default should be: *, 80, * where the first * is from server.bind and 80 is from server.port.

If both IPv4 and v6 are available, is there a reason to only bind to one of the two?

The reason is simple: I don't want to change it.

I understand. Do you have any idea why Debian uses $SERVER["socket"] == "[::]:80" instead of server.use-ipv6?

Olaf-van-der-Spek wrote:

I understand. Do you have any idea why Debian uses $SERVER["socket"] == "[::]:80" instead of server.use-ipv6?

Probably so it doesn't use v4-mapped addresses (::ffff:1.2.3.4) in REMOTE_ADDR, logs etc.

andersman wrote:

Probably so it doesn't use v4-mapped addresses (::ffff:1.2.3.4) in REMOTE_ADDR, logs etc.

Hasn't that been solved already by Lighttpd itself?

No, lighty still prints the address as it gets it from the socket (but debian disabled dual-stack, see below).

server.use-ipv6 = "enable" doesn't do much; it only uses
a) ipv6 [::] if you didn't specify server.bind
b) tries to use ipv6 for dns

If you use an ipv6 address in server.bind or $SERVER["socket"] lighty detects ipv6 itself, so no need for the option in these cases.

And i guess it fails if you try server.use-ipv6 with an ipv4 address.

Since debian changed the kernel default for bind-v6only to 1, server.use-ipv6 would lead to ipv6-only listening;
otoh, $SERVER["socket"] == "[::]:80" fails for people who either changed server.port or changed bind-v6only back to 0 (or tried backporting it).

So server.bind = [::] might be better. Or would that disable IPv4 (if v6only is enabled)?

What about BSD? It has v6only enabled too AFAIK.

fbsd set this in their config if you enable ipv6 when you are going to compile:
server.use-ipv6 = "enable"
$SERVER["socket"] == "0.0.0.0:80" { }

stbuehler wrote:

Do we listen to "[::]" now or "[::1]" ?

Neither. 127.0.0.1 is a v4 address, so only v4 gets bound.
If server.bind = "localhost", both v4 and v6 would get bound.

Would it be possible to support $SERVER["socket"] == "[::]" (just IPA, no port) and take the port from server.port? It's backwards compatible and solves part of the problem.

• detect unix domain socket path earlier
• detect IPv6 addr without port (might contain ':' within addr, e.g. [::])

After the above, Olaf's request becomes a smaller change to consider to use the default port (srv->srvconf.port) when one is not provided. This specifically errors if string is empty (""), with no host and no port. We'd have to consider how it might affect condition matching in configfile-glue.c were it to be allowed.

diff --git a/src/network.c b/src/network.c
index b401787..12c515b 100644
--- a/src/network.c
+++ b/src/network.c
@@ -205,11 +205,14 @@ static int network_server_init(server *srv, buffer *host_token, specific_config
                size_t len = buffer_string_length(b);
                char *sp = NULL;
                if ((b->ptr[0] == '[' && b->ptr[len-1] == ']') || NULL == (sp = strrchr(b->ptr, ':'))) {
-                       if (NULL == sp) {
-                               log_error_write(srv, __FILE__, __LINE__, "sb", "value of $SERVER[\"socket\"] has to be \"ip:port\".", b);
+                       if (*b->ptr == '\0') {
+                               log_error_write(srv, __FILE__, __LINE__, "s", "value of $SERVER[\"socket\"] has to be \"ip:port\"; value may not be empty");

                                goto error_free_socket;
                        }
+                       /* use server.port if set in config, or else default from config_set_defaults() */
+                       port = srv->srvconf.port;
+                       sp = b->ptr + len; /* point to '\0' at end of string */
                }

                /* check for [ and ] */

seem this changeset breaks this type of config:

# listen to ipv6
$SERVER["socket"] == "[::]:80" {  }

Thanks for the report. I have committed a patch to address the issue and it will push to master on Monday.

roytam1: the issue you reported should be fixed with current tip of git master. Thanks.

gstrauss wrote:

roytam1: the issue you reported should be fixed with current tip of git master. Thanks.

Fix confirmed, Thanks!