On 01/18/22 18:19, Eric Blake wrote:
On Tue, Jan 18, 2022 at 02:48:33PM +0100, Laszlo Ersek wrote:
The AI_ADDRCONFIG hint of getaddrinfo() is supposed to restrict the name resolution to such address families (IPv4 vs. IPv6) for which the resolving host has publicly routable addresses assigned.
The main problem with AI_ADDRCONFIG can be shown with the following command line:
$ nbdkit -f -p 32776 -P pidfile -i ::1 -exit-with-parent null
s/-exit/--exit/
Ah yes. TBH I noticed this myself yesterday, but not in the commit message -- in the terminal where I copied & pasted the command line from. There must have been something up with UTF-8 and the clipboard or whatever, because the terminal kept somehow mangling the double-dash at the front of "--exit-with-parent". So I fixed that ultimately in the terminal, but not in the commit message.
On a host where ::1 is the only IPv6 address assigned (namely to the loopback interface), the command fails with
nbdkit: getaddrinfo: ::1: 32776: Address family for hostname not supported
due to the "publicly routable" requirement.
Remove AI_ADDRCONFIG from the getaddrinfo() hints, and as a replacement, introduce the "-4" and "-6" options, similarly to netcat and ssh.
Hmm, I said in my earlier reply that it may not be worth the extra effort, but now that you've done it, it would be a shame to rip it out.
Thanks, I feel the same way! :)
(1) This makes options of the form:
-i 127.0.0.1 -i ::1
work regardless of "public" IPv6 / IPv4 connectivity;
(2) options of the form
-i localhost -i FQDN
will bind both IPv4 and IPv6 addresses of the desired interface(s);
(3) omitting the option "-i" will bind both IPv4 and IPv6 wildcard addresses (0.0.0.0 and ::);
(4) the configurations in (2) and (3) can be restricted to IPv4 or IPv6 addresses by adding the "-4" or "-6" option, respectively.
Importantly, this change allows the "connect-tcp6" test case of libnbd to pass on such hosts that have no IPv6 connectivity (i.e., where the only assigned IPv6 address is ::1, namely on the loopback interface).
Ref: https://listman.redhat.com/archives/libguestfs/2022-January/msg00110.html Signed-off-by: Laszlo Ersek <lersek@redhat.com> ---
+++ b/server/main.c @@ -86,6 +86,7 @@ static void error_if_stdio_closed (void); static void switch_stdio (void); static void winsock_init (void);
+int tcpip_sock_af = AF_UNSPEC; /* -4, -6 */ struct debug_flag *debug_flags; /* -D */ bool exit_with_parent; /* --exit-with-parent */ const char *export_name; /* -e */ @@ -367,6 +368,14 @@ main (int argc, char *argv[]) exit (EXIT_FAILURE); #endif
+ case '4': + tcpip_sock_af = AF_INET; + break; + + case '6': + tcpip_sock_af = AF_INET6; + break;
Thus, if the user uses nbdkit -46 (or -64), the last one specified silently overrides the earlier one, rather than being diagnosed as conflicting or explicitly permitting both. The override effect matches the long-option naming --ipv4-only, so I can live with it, but I'm also open to the idea of explicitly adding code to diagnose both options at the same time as an error, if we think that's friendlier than silent override.
+++ b/docs/nbdkit.pod @@ -173,6 +173,24 @@ Display information about nbdkit or a specific plugin:
Display brief command line usage information and exit.
+=item B<-4> + +=item B<--ipv4-only> + +=item B<-6> + +=item B<--ipv6-only> + +When a non-numeric argument is passed to the I<-i> option (such as a +Fully Qualified Domain Name, or a host name from C</etc/hosts>), +restrict the name resolution to IPv4 or IPv6 addresses. + +When the I<-i> option is omitted, listen on only the IPv4 or IPv6 +address of all interfaces (C<0.0.0.0> or C<::>, respectively).
This effect (when -i is not used) is good justification for having -4/-6 support.
+ +When both I<-4> and I<-6> options are present on the command line, the +last one takes effect.
Ah, so you documented the override as intentional.
Yes, I had written some getopt() loops in the distant past, and remembered this aspect.
It is possible to use a git orderfile to create patches with documentation changes appearing first in the output (see scripts/git.orderfile), which can make it easier to review related changes in logical (rather than alphabetical filename) order.
Absolutely! My top-level $HOME/.gitconfig actually sets "diff.orderFile" to an order file I created as my "global default"; it's just that (a) this file does not deal with *.pod files at all yet (I don't think I've edited a POD file before), plus (b) I didn't know about "scripts/git.orderfile" in nbdkit, and so I had not added a "diff.orderFile" override to .git/config in my clone.
So with that, you have my ACK for the series, after addressing the minor typo I spotted.
Thank you! Laszlo