We just provide the synchronization point, it's up to services that manages network to order some of their wait-online before that and make sure that the network is really up.
I assume that you are using NM, because the network-scripts are totally synchronous and their are ordered even before network.target.

Could you attach a debug-logfile or NetworkManager?


Enable debug-logging by configuring

[logging]
level=DEBUG

in /etc/NetworkManager/NetworkManager.conf


Afterwards, reboot the machine, ensure the issue got successfully reproduced and show the output of `journalctl -u NetworkManager -b0`



Thank you!!

Issue seems to be not related to NetworkManager. It's due to IPv6 DAD, you can find similar bugs like really old but still not fixed #782042.

Note that if you're not setting ipv6.may-fail=no already, you'll want to do that.  That ensures that NM will block startup for IPv6 to complete on that specific connection.

nmcli connection modify "<name of connection or UUID>" ipv6.may-fail no

(In reply to Dan Williams from comment #5)
> Note that if you're not setting ipv6.may-fail=no already, you'll want to do
> that.  That ensures that NM will block startup for IPv6 to complete on that
> specific connection.
> 
> nmcli connection modify "<name of connection or UUID>" ipv6.may-fail no

This is required, but probably not sufficient, as we signal the
activation before IPv6 DAD terminates when addresses are still
tentative and so a bind() to a static address will fail:

 $ nmcli c add type ethernet ifname veth0 con-name veth0+ autoconnect no ip4 1.2.3.4/32 ip6 fd01::42/64
 $ nmcli c mod veth0+ ipv6.may-fail no

 $ nmcli connection up veth0+ && socat -d -d TCP6-LISTEN:1234,bind=[fd01::42] UNIX-CLIENT:/tmp/foo.sock
  Connection successfully activated (D-Bus active path: /org/freedesktop/NetworkManager/ActiveConnection/10)
  socat[18988] E bind(3, {AF=10 [fd01:0000:0000:0000:0000:0000:0000:0042]:1234}, 28): Cannot assign requested address
 $ ip a show veth0
 [...] inet6 fd01::42/64 scope global tentative [...]

 $ ip a show veth0
 [...] inet6 fd01::42/64 scope global [...]
 $ socat -d -d TCP6-LISTEN:1234,bind=[fd01::42] UNIX-CLIENT:/tmp/foo.sock
  socat[19443] N listening on AF=10 [fd01:0000:0000:0000:0000:0000:0000:0042]:1234

This is not consistent with IPv4 since we guarantee that if
ipv4.may-fail=no the IPv4 addresses will be usable immediately after
activation (of course, since duplicate detection is done in userspace
before assigning addresses to interface).

How about branch bg/ipv6-dad-rh1243958 ? AFAIU systemd-networkd does
something similar [1].

Maybe we should also add support for configuring the DAD timeout for
IPv6 (IPv4 already has it).

[1] https://github.com/systemd/systemd/issues/650

This bug was accidentally moved from POST to MODIFIED via an error in automation, please see mmccune with any questions

(In reply to Beniamino Galvani from comment #6)

> How about branch bg/ipv6-dad-rh1243958 ? AFAIU systemd-networkd does
> something similar [1].

dad6_timeout() always fails the connection on timeout. Don't we have a check-fail, which only fails when may-fail=no?


Also, can you clear dad6_timer_id before the state-change in dad6_timeout()?


priv->dad6_ip6_config and priv->dad6_timer_id is set in activate_stage5_ip6_config_commit(), but there are few places where those fields are cleared. If you are sure that you don't overwrite a previous setting, add a

   g_assert (!priv->dad6_ip6_config);
   g_assert (!priv->dad6_timer_id);
   priv->dad6_ip6_config = nm_ip6_config_new (nm_device_get_ip_ifindex (self));
   ...


It general, it seems that after clearing dad6_timer_id, you can always clear dad6_ip6_config too.

(In reply to Thomas Haller from comment #8)
> dad6_timeout() always fails the connection on timeout. Don't we have a
> check-fail, which only fails when may-fail=no?

We have a check_ip_failed(), but it doesn't fail immediately when may-fail=no. The correct fix for this is in bug 741347.

> Also, can you clear dad6_timer_id before the state-change in dad6_timeout()?

Ok.

> priv->dad6_ip6_config and priv->dad6_timer_id is set in
> activate_stage5_ip6_config_commit(), but there are few places where those
> fields are cleared.

Yes, it turns out that the fields could be overwritten, I added a check.

Branch repushed.

if (nm_logging_enabled (LOGL_TRACE, LOGD_DEVICE | LOGD_IP6)) {
    ...
    _LOGt (...)
}

should be

if (_LOGt_enabled (LOGD_DEVICE | LOGD_IP6)) {
    ...
}


The reason is, that _LOGt() is a NOP unless compiled with --with-more-logging. Thus, you also want to use _LOGt_enabled() which returns always FALSE in that case.



https://bugzilla.gnome.org/show_bug.cgi?id=741347 seems important, and should be fixed.

(In reply to Thomas Haller from comment #10)
> The reason is, that _LOGt() is a NOP unless compiled with
> --with-more-logging. Thus, you also want to use _LOGt_enabled() which
> returns always FALSE in that case.

Fixed.

> https://bugzilla.gnome.org/show_bug.cgi?id=741347 seems important, and
> should be fixed.

It has been merged now and I've updated bg/ipv6-dad-rh1243958 to use the new function for failing an IP method.

Please add more comments there -- describe what ipv6_dad_start() does. It's not immediately obvious to me why would it start dad when kernel does that.

Also, perhaps we could add an option to use optimistic DAD; or just honor what the sysctl is set to? (If it even is a good idea; maybe it isn't)

(In reply to Lubomir Rintel from comment #12)
> Please add more comments there -- describe what ipv6_dad_start() does. It's
> not immediately obvious to me why would it start dad when kernel does that.

I did some changes:
 - reworded commit message and comments
 - removed timer
 - better deal with DAD failures
 - added ipv6_get_tentative_addresses()

and repushed.

> Also, perhaps we could add an option to use optimistic DAD; or just honor
> what the sysctl is set to? (If it even is a good idea; maybe it isn't)

It seems that the optimistic flag can only be added by kernel, and to autoconfigured addresses. Since NM configures addresses by itself I don't think optimistic DAD would be possible, if I understand well how it works...

+         tmp_config = nm_ip6_config_new_cloned (priv->ext_ip6_config_captured);
+         nm_ip6_config_intersect (tmp_config, priv->dad6_ip6_config);
+         num = nm_ip6_config_get_num_addresses (tmp_config);
+
+         for (i = 0; i < num; i++) {
+              addr = nm_ip6_config_get_address (tmp_config, i);
+              dad_pending |=    (addr->n_ifa_flags & IFA_F_TENTATIVE)
+                             && !(addr->n_ifa_flags & IFA_F_DADFAILED);
+         }

maybe you could add a function

  gboolean
  nm_ip6_config_has_any_dad_pending (const NMIP6Config *self,
                                     const NMIP6Config *candidates)

which basically iterates over the addresses of @self (ignoring addresses
which are not also in @candidates), and returns TRUE if there are any addresses that are TENTATIVE.

it allows a more efficient implementation, which doesn't involve cloning+intersecting the NMIP6Config instance.


Interesting, it uses nm_platform_ip6_address_get(). I didn't expect this function to be ever useful beside testing :)


s/nm_platform_get ()/NM_PLATFORM_GET/





priv->dad6_ip6_config = ipv6_get_tentative_addresses (self);

these two items are closely related, but their name is completely different. I think they should be named somehow similar (I like same prefixes)...
for example:
  priv->dad6_ip6_config = dad6_get_tentative_addresses (self);
  priv->dad6_ip6_config = dad6_ip6_config_get (self);




in ipv6_get_tentative_addresses(), could we cache nm_device_get_ip_ifindex() in a variable ifindex? Possibly with

  ifindex = nm_device_get_ip_ifindex();
  g_return_val_if_fail (ifindex > 0, NULL);





And can we construct dad6_config lazy, when needed:

     if (pl_addr && (pl_addr->n_ifa_flags & IFA_F_TENTATIVE)) {
         if (!dad6_config)
              dad6_config = nm_ip6_config_new (nm_device_get_ip_ifindex (self));

I wouldn't expect that finding dad-candidates happens in most of the cases.





+         /* Check if we have to wait for DAD */
+         if (priv->ip6_state == IP_CONF && !priv->dad6_ip6_config) {

doesn't that mean, we only create the dad6_ip6_config instance once? What if there are multiple calls to ip6_config_merge_and_apply(), don't we want to re-generate it every time?



(I like the commit message, and the patch in general)

(In reply to Thomas Haller from comment #14)

> maybe you could add a function
> 
>   gboolean
>   nm_ip6_config_has_any_dad_pending (const NMIP6Config *self,
>                                      const NMIP6Config *candidates)
> 

> s/nm_platform_get ()/NM_PLATFORM_GET/

> these two items are closely related, but their name is completely different.
> I think they should be named somehow similar (I like same prefixes)...

> in ipv6_get_tentative_addresses(), could we cache nm_device_get_ip_ifindex()
> in a variable ifindex? 

> And can we construct dad6_config lazy, when needed:

Fixed all the above and repushed.


> +         /* Check if we have to wait for DAD */
> +         if (priv->ip6_state == IP_CONF && !priv->dad6_ip6_config) {
> 
> doesn't that mean, we only create the dad6_ip6_config instance once? What if
> there are multiple calls to ip6_config_merge_and_apply(), don't we want to
> re-generate it every time?

activate_stage5_ip6_config_commit() can be called multiple times but only one succeeds and changes the state from IP_CONFIG to IP_CHECK. When that state transition is about to happen we are guaranteed to have all necessary addresses (static, SLAAC or DHCP for managed conf), so I think it is safe to ignore, from a DAD perspective, any following call.

For me it looks good now.

Merged to master:

https://cgit.freedesktop.org/NetworkManager/NetworkManager/commit/?id=7dbabe9bdca2d9c3b9df11bea52cdd500f699d31

nm-online exits when tentative addresses are gone (dad check was performed before), that is different to 1.0.6 behaviour.

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://rhn.redhat.com/errata/RHSA-2016-2581.html