Instead of accessing the singleton getter nm_settings_get(), obtain
the settings instance from the device instance itself via
nm_device_get_settings().
Currently NM proceeds with the activation of a device just after the
IPv6 configuration is applied. Server applications will bind to IPv6
addresses as soon as NM signals the presence of network connectivity,
but since the addresses are still tentative the bind will fail. There
are a couple of solutions to this.
Linux kernel supports "optimistic DAD", which is a modification of
Neighbor Discovery and SLAAC processes that allows addresses to be
used (under certain contraints) while kernel is performing DAD on
them. However it is not feasible to let NM enable optimistic DAD for
the devices it controls for the following reasons:
- it is not guaranteed to be always available since it can be turned
off at compile time
- RFC 4429 states that it should not be used for manually entered
addresses
- it works only with autoconf addresses generated by kernel
Therefore, use a different approach and handle this in NM by waiting
that the kernel completes DAD before continuing activation. We build a
list of addresses that are tentative just after the new configuration
is applied and then we asynchronously wait a platform address-change
event where all NM-configured addresses become non-tentative.
A similar solution has been adopted also by other network managing
tools:
https://anonscm.debian.org/cgit/collab-maint/ifupdown.git/commit/?id=ec357a5d6cb5fa8b0004c727d7cc48253c59eb0f8012cd3919https://bugzilla.redhat.com/show_bug.cgi?id=1243958
Fall back to system default value for ipvx.dns-priority when it's zero
in the setting. For VPNs the default value is 50; for other
connections is 100, but it depends also on the content of
[connection*] sections in NetworkManager.conf.
This is especially important, because changing MTU takes the
link down for a moment. Taking a link down deletes IP routes and
IPv6 addresses. Thus, when the link comes up again, we must restore
them.
Otherwise, we don't call merge_and_apply() until the next DHCP lease
(or possibly never in case of static addressing).
https://bugzilla.redhat.com/show_bug.cgi?id=1309899
nm_device_set_ip4_config() is called during cleanup and
from ip4_config_merge_and_apply(). The latter, has several
call sites.
It's not easy to track whether we called set_ip4_config with
or without commit (and if we call it without commit, we might
not see a logging line at all).
(same for nm_device_set_ip6_config()/ip6_config_merge_and_apply()).
We would subscribe to config-changed signal during object-realize,
however only unsubscribe during dispose().
Avoid multiple subscributions, and unsubscribe also when unrealizing
the device.
Also, always subscribe to the signal, even without capability
NM_DEVICE_CAP_CARRIER_DETECT. In the next commit, we will re-read
capabilities later on, so just always subscribe.
Contrary to gboolean, bool is only one byte in size.
Due to alignment and ordering of the fields, this saves
merely 16 bytes per NMDevicePrivate struct (on x86_64),
still.
Also, bool is coerced by the compiler to be strictly FALSE or
TRUE -- contrary to gboolean, which can be any integer.
Thus, for bool type, "g_assert (NM_IN_SET (value, FALSE, TRUE));"
never fails. That is desirable as well.
While not a large win, it seems favorable to use bool type for
fields of a struct.
When the IP status is IP_DONE and a DHCP transaction succeeds the
'dhcp4' and 'dhcp6' pending actions must be removed. Without this, a
temporary link loss just after the activation would cause a DHCP
restart and those actions would remain set, blocking the startup.
https://bugzilla.redhat.com/show_bug.cgi?id=1330893
Once we start with dispose, we certainly don't want to process any platform
events for the device anymore.
Previously, we disconnect those handlers only later during dispose, so it's
not clear that we would not receive a device_ipx_changed signal after _cleanup_generic_pre().
Fix this possible (or actual) bug.
Since commit a47c13a7a2, update_ip4_config() re-schedules
itself in case activate_stage5_ip4_config_commit is pending. Thus, there is no need to
cancel any queued queued_ip4_config_id.
Also as that does not properly fix the issue unlike a47c13a7a.
update_ip4_config() and update_ip6_config() are called from nm_device_capture_initial_config().
At that point, we don't expect any activation-source scheduled, thus the "if" should not
not be hit anyway.
So, this patch should actually make no difference, but it seems clearer
to me. Also, because it would be a bug to re-schedule the idle handler
that is already pending, but from inspecting nm_device_capture_initial_config()
it is not immediately clear that this cannot be the case.
Make DHCPv6 more robust WRT temporary failures of servers by retrying
DHCP for a predefined number of times at regular intervals when the
lease expires.
https://bugzilla.gnome.org/show_bug.cgi?id=741347
Make DHCPv4 more robust WRT temporary failures of servers by retrying
DHCP for a predefined number of times at regular intervals when the
lease expires.
https://bugzilla.gnome.org/show_bug.cgi?id=741347
Introduce the nm_device_ip_method_failed() function to check if the
failure of an IP method should cause the activation to fail, and use
it where appropriate.
http://bugzilla.gnome.org/show_bug.cgi?id=741347
When a new dynamic configuration is received, it is stored in a member
of private structure (e.g. @dhcp6_ip6_config) and a commit is
scheduled. Before the commit is executed, an update_ipx_config() could
be called and it would change the configuration before it is
committed.
This race condition causes failures in assigning the addresses
received through DHCPv6 when the internal client is used (but
potentially other clients and methods are affected).
To fix it, postpone updates of IP configurations when a commit is
already pending.
The "source" field of NMPlatformIPRoute (now "rt_source") maps to the
protocol field of the route. The source of NMPlatformIPAddress (now
"addr_source") has no direct equivalent in the kernel.
As their use is different, they should have different names. Also,
the name "source" is used all over the place. Hence give the fields
a more distinct name.
Already previously, the mode and rc-manager were intertwined in a complicated
way:
- dns=none effectively disables rc-manager.
- if resolv.conf was immutable, it would disable the rc-manager
by setting "resolv_conf_mode=NM_DNS_MANAGER_RESOLV_CONF_UNMANAGED".
- resolv_conf_mode was anyway a redundant piece of information to
rc_manager.
Now there are only two relevant settings: priv->plugin and
priv->rc_manager. And they can be set independently from each other.
Before that was not possible. For example, you could not set a
dns plugin with rc-manager=unmanaged (the only way to achive that
was via an immutable resolv.conf or by having rc-manager=symlink
and let resolv.conf link somewhere else.
The infiniband drivers don't implement the rtnetlink link deletions.
Therefore we unrealize the NMDevice instance but the backing resources
stay around, preventing us from ever realizing the device again.
When we want to preserve the default-route on cleanup, we must first
set it to assumed, before clearing it. Otherwise, NMDefaultRouteManager's
update() will delete the default route.
This is the oposite of the deconfigure case, where we first set it to
!has && !assumed, to force the route-manager to delete the route.
Add a function _update_default_route() to set the default_route
flags and call update() in one step.
Also, if there are no changes, skip the call to NMDefaultRouteManager's
update().
