The SpinBox value returned by get_value and friends gets definitely
updated only when the control loses focus. That is unfortunately too late
for the validation machinery, so force the updates at the right time.
Part-of: <https://gitlab.gnome.org/GNOME/gnome-control-center/-/merge_requests/1983>
When the IP method is "disabled" or "shared" then everything else is
supposed to be insensitive. This currently fails if you toggle between
the two, because it's implemented using property bindings that are just
not smart enough to handle this task. Handle sensitivity only in
method_changed() to avoid this.
Additionally, not all of the widgets are being consistently
disabled/enabled when appropriate. E.g. when the method is "local" then
only the DNS entry, route entries, and default route checkboxes become
insensitive, leaving the other widgets, including notably the Automatic
switches, sensitive. They should all become insensitive, as when the
method is "disabled" or "shared." Fix this by organizing all the related
widgets into boxes and setting the sensitivity of the entire box. (Note
the strategy followed here does not exactly match nm-connection-editor,
which always allows editing addresses. We only allow that in Manual
mode. I'm not sure if this is advisable or not, so won't touch that.)
Finally, the Automatic DNS and Automatic Routes toggles should only be
sensitive when the method is "Automatic".
When the entry is initially empty, there is no error. If you enter
anything and then delete it, the entry is left in the error state. Empty
should not be an error.
This reverts commit 8a4a80b7b2.
This is a manual revert, because the code changed considerably in
a2b9620b1b. Anyway, although this seemed
like a good idea, problem is it clobbers the original state of the
connection without any explicit user action if the connection is
configured to use both manual and auto DNS. And this happens in practice
for imported Wireguard connections, which uselessly have auto DNS
enabled due to this NetworkManager bug:
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/1399
Additional discussion:
https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1745#note_2112420
The simplest solution is to not try to prevent the user from configuring
both manual and auto DNS. Instead, let's just warn the user that this
configuration may not be intended in a follow-up commit.
Fixes#2617
Even though the routes_metric_label is in a GtkSizeGroup with the
GtkEntry for the metric, its size was set too big after adding the entry
to the size group. To fix this, add all the other labels and
corresponding entries to size groups as well. The hexpand can then be
removed as well on the labels.
Fixes#1235
For example, fix adding new VPN connections.
In 60b4956c05 I correctly observed that we
need to not run code that requires a device when there is no device.
NetConnectionEditor is a multipurpose dialog and self->device is
optional when creating the dialog. E.g. when modifying VPN
configuration, we update just the configuration, not an NMDevice.
However, I added this check too soon, before updating the connection
configuration. We need to update the configuration first, then only bail
before proceeding to update the device, not sooner.
Fix#2668
If the NMDevice is not active (i.e. if we are editing a connection that
is not active) then don't warn when failing to reapply changes to the
device. That's expected and not something we should warn about.
We could perhaps not even try, but the device could become inactive
between the time of our check and this error, so better ignore the error
regardless.
All of the following code assumes that self->device is valid, so we need
to skip over it. It's confusing, but this is a multipurpose dialog and
self->device is optional when creating the dialog. E.g. when modifying
VPN configuration, we update just the configuration, not an NMDevice.
If the operation is cancelled (because the dialog was closed, because
the Apply button was pressed), then trying to make further use of the
source_object is a use after free, which is bad. At first I tried to fix
this by simply avoiding the use after free when the operation is
canceled, but then I realized it is ridiculous to always try committing
connection changes when closing the dialog, then immediately cancel the
operation by destroying the dialog.
So instead I've decided to not pass the cancellable along to these
operations, and instead ref the dialog to keep it alive until the
operations complete. Instead, let's just hide the window.
This commit also removes an inaccurate comment /* Leave the editor open
*/ placed right before the call to the function that hides the editor.
There's no need to leave the editor open when updating the device fails.
The connection properties at least are still saved.
Fixes#2618
It required two changes:
* Setting the stack page only after all pages are initialized
* And removing the VPN choices list from its bin only after that
Hooking to all the toggled signals from all the buttons for executing
the same action is inneficient, and can potenticall end up in a segmentation
fault due to some race in the signal emmission, where the active button
gets deactivated before the clicked button is activated
Looking at the GTK4 code, in a radio group:
- The button which was previously active gets de-activated, emitting its
corresponding toggled signal.
- The active property for the clicked button gets set.
- The clicked button emits its toggled signal.
Therefore, if the first toggle signal gets processed before the active
property is set, there can be a race condition. We are seeing this downstream
at pmOS: https://gitlab.com/postmarketOS/pmaports/-/issues/1816
Instead of this racy behavior, follow upstream recommendation and keep track
of the state through a stateful signal.
Meson extracts them by itself and add them as dependencies for the target.
It means one less location to keep track of files, and a lot less boilerplate
around the meson files
Destructive actions are supposed to have a dialog asking the user for confirmation.
This commit adds the confirmation dialog when user tries to forget a wifi network.
Fixes#2371
The maximum MTU value of 10000 is too low for USB Ethernet, which has a
maximum (for Linux USB gadgets) of 15412 bytes (although the upper limit
is the USB wMaxPacketSize which goes up to 4294967295 bytes):
linux/drivers/usb/gadget/function/u_ether.c:#define GETHER_MAX_MTU_SIZE 15412
Multiple Intel NICs can use an MTU of 16110 bytes:
linux/drivers/net/ethernet/intel/e1000/e1000_hw.h:#define MAX_JUMBO_FRAME_SIZE 0x3F00
linux/drivers/net/ethernet/intel/e1000e/defines.h:#define MAX_JUMBO_FRAME_SIZE 0x3F00
linux/drivers/net/ethernet/intel/igbvf/defines.h:#define MAX_JUMBO_FRAME_SIZE 0x3F00
The NetworkManager limit is 4294967295 bytes but this is unreasonable
in a typical enivornment because of the memory required for packets of
that size.
The maximum IPv4 and IPv6 (without using Jumbograms) packet size is 65535
bytes so increase the maximum MTU value to 65536 allow full size IP
packets to be used.
There is a corresponding change in network-manager-applet.
When NetworkManager doesn't give us any secrets for a connection, the
connection editor still needs to be opened. This change ensures that
initialization of the editor completes even when there's an error when
fetching secrets.
Fixes#2329
GtkCheckButton accepts a widget as its child and will toggle when
clicking its child, currently we put check button and label of metered
connection into a box, then clicking the label won't toggle the check
button, this differs from the other two check buttons.
This commit makes the label a child widget of the check button, so the
three check buttons behave the same.
When the dialog is closed using ESC key, the "close-request" signal is
emitted in addition to the "response" signal. When the "close-request"
is handled, it frees the memory to which info points. In the "response"
signal handler, the memory of info pointer is accessed again, leading
to a segmentation fault.
Fix this by removing the "close-request" function callback, which shares
the same behaviour as the "response" callback function.
Fixes: https://gitlab.gnome.org/GNOME/gnome-control-center/-/issues/2320
