Let the link-add functions return the internal pointer to the platform
link object. Similar to link-get, which doesn't copy the link either.
Also adjust the sole users of the add-functions (create-and-realize)
to take the pointer.
Eventually we still copy the returned data, because accessing platform can
invalidate the returned pointer. Thus we don't actually safe any copying
by this (at least every use of the function currently leads to the data
being copied).
Still change it, because I think the API of NMPlatform should look like that.
NMPlatformLink is a plain struct (not a GObject, for which we usually
don't use const). We certainly don't want the functions to modify the
passed-in data.
There are the link-types NONE and UNKNOWN. NONE is a linktype that is never
returned by platform, but UNKNOWN is very much a valid (albeit unspecified)
type.
Effectively, create_and_realized() should create a link of a known type,
thus it should never return an UNKNOWN link type at this point. Still
change it because it feels more correct.
Device subclasses (for example NMDeviceWifi) can use the ifindex in
their constructor(), but the value now is set later in
parent class constructed(). This causes the following:
nm_platform_wifi_get_capabilities: assertion 'ifindex > 0' failed
Fix this by initializing ifindex earlier in NMDevice's
constructor(). While at it, remove the
nm_assert (pllink->type != NM_LINK_TYPE_NONE);
assertion, since pllink can be NULL there.
Fixes: 6db04dc206
This is to make it possible for the device factories to indicate the desired
link type and make it possible to avoid matching the unrealized device to a
platform device of different link type.
We often lookup the private data and retrieve it via NM_DEVICE_GET_PRIVATE(),
which in turn calls G_TYPE_INSTANCE_GET_PRIVATE().
Instead cache the pointer to the private data.
There are up- and downsides:
- requries additional sizeof(gpointer) bytes for each NMDevice.
+ retrieving the private pointer will be slightly faster.
+ easier debugging in gdb as it is currently often a pain to
retrieve the private data.
But most importantly, the allows to change our common pattern
to first cache the private data in a variable @priv. That is
often cumbersome to write, especially for short functions.
This change gives us a choice to use self->priv directly.
Such a change should not be aimed for every class. Instead it makes
mostly sense for NMDevice, where it pays off better due to the
class' size and ubiquitous use.
https://mail.gnome.org/archives/networkmanager-list/2015-December/msg00017.html
Showcase for the new macros NM_UTILS_STRING_LOOKUP_TABLE() and
NM_UTILS_STRING_LOOKUP_TABLE_DEFINE_STATIC().
It changes behavior in case of looking up an invalid/unknown
state reason. Previously it would just have returned "unknown"
-- which was indistinguishable from a regular "unknown" value.
Now it returns the numeric id as a string. The string is allocated
with alloca(), which is desired but one should be aware of the pitfalls:
- prevents the caller from being inlined
- bad idea to do in a loop.
I found the handling of the master-device very confusing because it was
unclear who sets priv->master, and when it should be set.
Now:
- Setting priv->master (in a slave) always goes together with adding
the master to priv->slaves (in the master). Previously, this was
done at separate places, so it was not clear if master and slave
always agree on their relationship -- in fact, they did not.
- There are now three basic functions which do the enslaving/releasing:
(1) nm_device_master_add_slave()
(2) nm_device_master_enslave_slave()
(3) nm_device_master_release_one_slave()
Step 3/release basically undoes the 1/add and 2/enslave steps.
- completing the enslaving/releasing is now done by
(1) nm_device_slave_notify_enslave()
(2) nm_device_slave_notify_release()
These functions also emit signals like NM_DEVICE_MASTER.
- Derived classes no longer emit NM_DEVICE_SLAVES notification. Instead
the notification is emited together with NM_DEVICE_MASTER, whenever a
slaves changes state. Also, NM_DEVICE_SLAVES list now only exposes
slaves that are actually @is_enslaved.
Instead of reimplementing the slave property in bond, bridge
and team, just add the property to the parent class. It's not
that the parent class would be agnostic to the master/slave
implementation, all the slaves are known to the every device
type implementation.
Also, the derived class doesn't know the correct time when
to invoke the notify-changed for the slaves property.
E.g. it should be only invoked after nm_device_slave_notify_enslave()
when other components also consider the slave as enslaved.
Later this will be fixed so that the SLAVES property correspond
to what other master/slave related properties say.
release_slave() should do the right thing and handle errors as
good as it can. There is no value in propagating the error and
it's not clear what the caller should do in face of a failure
during release.
We have the master/slave related functions
- for master device:
- nm_device_master_add_slave()
- nm_device_master_release_slaves()
- nm_device_release_one_slave()
- nm_device_enslave_slave()
- for slave device:
- nm_device_slave_notify_enslave()
- nm_device_slave_notify_release()
Rename the two that didn't match the pattern to
- nm_device_master_release_one_slave()
- nm_device_master_enslave_slave()
But, of course, only one realized device can have the same
interface name at a time.
This commit effectively reverts most of:
1b37cd0340
core: allow ActiveConnections to be created without a device
But it's not easy to do a separate revert of that code due to
interdependencies with nm-manager.c.
Creating devices when they are defined by a connection also makes
makes it possible to require the NMDevice to be present when
activating it, which means we can remove a bunch of code from
NMManager that had to handle software devices not existing yet at
the time of the activation request.
But it also means we must be more careful when finding master
interfaces during slave activation, since we cannot simply match
by interface name alone. Instead we must find the master which
matches both the interface name and can control slaves of the type
which is being activated.
Ensure the platform link with the same interface name as the
NMDevice is actually compatible with it before using the link
for initialization of device properties. If not, remove the
NMDevice and create a new one since there are kernel resources
with a different type.
Unrealized devices aren't backed by kernel resources and so won't know
all of their attributes. That means three things:
1) they must update their attributes when they become realized
2) they must clear those attributes when unrealized
3) they must be looser in checking compatible connections until
they are realized
This requires that the setup() function be split into two parts, start & finish,
because finish must be run after add_device()
Also, we can simplify whether to pay attention to 'recheck-assume', which
is now dependent on priv->is_nm_owned, because the only case where NM should
*not* listen for the 'recheck-assume' signal is when the device is a
software device created by NM itself. That logic was previously spread
across the callers of add_device() but is now consolidated into
nm-manager.c::device_realized() and nm-device.c::nm_device_create_and_realize().
This property is TRUE for devices that exist either as a kernel device
or are backed by some other resource (eg, ModemManager object, Bluez
device, etc). It will eventually be FALSE for software devices that
are not yet instantiated.
There is only one user of the NM_DEVICE_MASTER property (NMActiveConnection).
device_master_changed() uses the property for change notifications,
but returns early if !nm_device_get_master().
We might emit the "notify::master" signal too often, even when nothing
changed (because currently we always emit it when priv->master changes).
If we would fix that to only emit the signal when nm_device_get_master()
actually changes, it would still be correct, because device_master_changed()
doesn't care about notifications when the master is unset.
Hence, the only user doesn't care about the difference. So change it because
it's unexpected that the property and function are not completely equal.
Commit cd3df12c8f reused the
virtual function component_added() to notify the vlan device
about a possibly new parent.
This reuse of the virtual function for another purpose is confusing.
Clean that up by splitting the implementation and add a new
virtual function nm_device_notify_new_device_added() which gets
(only implemented by NMDeviceVlan).
When deleting a master-device either via `nmcli device delete`
or `ip link delete`, the slave-device would hang.
This seems to be broken for a very long time already.
This is due to the following:
#0 0x00005555555f548c in nm_device_slave_notify_release (self=0x555555dc1300, reason=NM_DEVICE_STATE_REASON_NONE) at devices/nm-device.c:2175
#1 0x00005555555d6de2 in nm_device_release_one_slave (self=0x555555de3dd0, slave=0x555555dc1300, configure=0, reason=NM_DEVICE_STATE_REASON_NONE) at devices/nm-device.c:1117
#2 0x00005555555f02b7 in device_link_changed (self=0x555555dc1300) at devices/nm-device.c:1460
Previously, nm_device_slave_notify_release() being called with reason
"NONE" did not actually transition the device-state, thus keeping the
device wrongly in activated state.
There were two callers that passed configure=FALSE to nm_device_release_one_slave(),
(and thus reason=NONE to nm_device_slave_notify_release()):
- (1) device_link_changed():
nm_device_release_one_slave (priv->master, self, FALSE, /*wrong reason NONE*/);
- (2) nm_device_removed():
nm_device_release_one_slave (priv->master, self, FALSE, NM_DEVICE_STATE_REASON_REMOVED);
We should always change the device-state during nm_device_slave_notify_release()
regardless of the reason.
(2) was added by commit c83b40aca7, later
refined by commit 5dd48f7527. In a way
change the second fix to perform some of the configuration (but still
not unenslaving the device).
The new object type represents tunnels over IPv4 and IPv6.
We have a single setting type (NMSettingIPTunnel) for tunnels and it
can't be shared among different device factories. So we define also a
single device type for all tunnels.
This new object will also represent GRE tunnels, which before were
instantiated as NMDeviceGre and had a ".Device.Gre" D-Bus
interface. This commit introduces a change in behavior.
This enum was unused and meaningless because the platform signals
are emitted as a consequence of netlink messages. It is not clear
whether a netlink message was received due to an external event
or an internal action.
The DHCP client from new libsystemd-network requires a link-local IPv6
address to be passed to the library; add a new argument to
nm_dhcp_manager_start_ip6() and related functions.
When deconfiguring a device, we must also explicitly clear the
default-route -- unless the device was assumed.
This can easily reproduced by disconnecting the cable from the
wired connection that has the default rout. Prevously, the
default-route was not cleared and lingered around.
https://bugzilla.gnome.org/show_bug.cgi?id=757587
Device subclasses can call nm_device_recheck_available() at any time,
and the function would change the device's state to UNKNOWN in cases
where the device was available already. For WWAN devices, availability
is rechecked every time the modem state changes, resulting in:
NetworkManager[28919]: <info> (ttyUSB4): modem state changed, 'disabled' --> 'enabling' (reason: user-requested)
NetworkManager[28919]: <debug> [1445538582.116727] [devices/nm-device.c:2769] recheck_available(): [0x23bd710] (ttyUSB4): device is available, will transition to unknown
NetworkManager[28919]: <info> (ttyUSB4): modem state changed, 'enabling' --> 'searching' (reason: user-requested)
NetworkManager[28919]: <debug> [1445538582.776317] [devices/nm-device.c:2769] recheck_available(): [0x23bd710] (ttyUSB4): device is available, will transition to unknown
Previously most objects were implicitly unexported when they were
destroyed, but since refcounts may make the object live longer than
intended, we should explicitly unexport them when they should no
longer be present on the bus.
This means we can assume that objects will always be un-exported
already when they are destroyed, *except* when quitting where most
objects will live until exit because NM leaves interfaces up and
running on quit.
