Iterating hash tables gives an undefined order. Often we want to have
a stable order, for example when printing the content of a hash or
when converting it to a "a{sv}" variant.
How to achieve that best? I think we should only iterate the hash once,
and not require additional lookups. nm_utils_named_values_from_strdict()
achieves that by returning the key and the value together. Also, often
we only need the list for a short time, so we can avoid heap allocating
the list, if it is short enough. This works by allowing the caller to
provide a pre-allocated buffer (usually on the stack) and only as fallback
allocate a new list.
This solves a bug exposed by the following cmds:
$ nmcli c add type bond ifname bond0 con-name bond0
$ nmcli c modify bond0 +bond.options miimon=100
$ nmcli -f bond.options c show bond0
bond.options: mode=balance-rr
Here we just added the option 'miimon=100', but it doesn't get saved in
because nm_settings_connection_set_connection() which is responsible for
actually updating the connection compares the new connection with old
one and if and only if the 2 are different the update is carried out.
The bug is triggered because when comparing, if default values are taken into
account, then having 'miimon=100' or not having it it's essentially the
same for compare(). While this doesn't cause a bond to have a wrong
setting when activated it's wrong from a user experience point of view
and thus must be fixed.
When this patch is applied, the above
commands will give the following results:
$ nmcli c add type bond ifname bond0 con-name bond0
$ nmcli c modify bond0 +bond.options miimon=100
$ nmcli -f bond.options c show bond0
bond.options: mode=balance-rr,miimon=100
Fix unit tests and also add a new case covering this bug.
https://bugzilla.redhat.com/show_bug.cgi?id=1806549
Add 'nm_setting_bond_get_option_normalized()', the purpose of this API
is to retrieve a bond option normalized value which is the option that
NetworkManager will actually apply to the bond when activating the
connection, this takes into account default values for some options that
NM assumes.
For example, if you create a connection:
$ nmcli c add type bond con-name nm-bond ifname bond0 bond.options mode=0
Calling 'nm_setting_bond_get_option_normalized(s_bond, "miimon")' would
return "100" as even if not specified NetworkManager enables miimon for
bond connections.
Another example:
$ nmcli c add type bond con-name nm-bond ifname bond0 bond.options mode=0,arp_interval=100
Calling 'nm_setting_bond_get_option_normalized(s_bond, "miimon")' would
return NULL in this case because NetworkManager disables miimon if
'arp_interval' is set explicitly but 'miimon' is not.
Add '_nm_setting_bond_get_option_or_default()' and move all the custom
policies applied by NM for bond options in there.
One such example of a custom policy is to set 'miimon' to 0 (instead of its
default value of 100) if 'arp_interval' is explicitly enabled
and 'miimon' is not.
This means removing every piece of logic from
nm_setting_bond_add_option() which used to clear out 'arp_interval' and
'arp_ip_target' if 'miimon' was set or clear out 'miimon' along with
'downdelay', 'updelay' and 'miimon' if 'arp_interval' was set.
This behaviour is a bug since the kernel allow setting any combination
of this options for bonds and NetworkManager should not limit the user
to do so.
Also use 'set_bond_attr_or_default()' instead of 'set_bond_attr()' as
the former calls '_nm_setting_bond_get_option_or_default()' to implement
the right logic to retrieve bond options according to current bond
configuration.
Fix 'miimon' and 'arp_interval' validation, they can both be set indeed,
the kernel does not impose this limitation, nevertheless is sensible to
keep the defaults as previously (miimon=100, arp_interval=0).
Also add unit test.
Doing 'verify()' with options such as 'miimon' and 'num_grat_arp' set to
arbitrary values it's not consistent with what NetworkManager does to
bond options when activating the bond through 'apply_bonding_config()'
(at a later stage) because the said values do not
correspond to what the default values for those options are.
This leads to an inconsistency with the 'miimon' parameter for example,
where 'verify()' is done while assuming it's 0 if not set but its
default value is actually 100.
Fixes: 8775c25c33 ('libnm: verify bond option in defined order')
Just looking at the hashtable entry of 'updelay' and 'downdelay' options
is wrong, we have to inspect their values to check if they're
actually enabled or not.
Otherwise bond connections with valid settings will fail
when created:
$ nmcli c add type bond ifname bond99 bond.options miimon=0,updelay=0,mode=0
Error: Failed to add 'bond-bond99' connection: bond.options: 'updelay' option requires 'miimon' option to be set
Also add unit tests.
https://bugzilla.redhat.com/show_bug.cgi?id=1805184
Fixes: d595f7843e ('libnm: add libnm/libnm-core (part 1)')
verify() should validate options in a deterministic order, so that
the same profile (with same libnm version) gives the same failure
reason every time.
Hence, visit the options in sorted order, like we do for nm_setting_bond_get_option().
Internally, the options are tracked in a hash table and of undefined
sort order. However, nm_setting_bond_get_option() always returns a stable
(sorted) order.
Move "mode" as first, because that is usually the most interesting option.
The effect is:
$ nmcli -o connection show "$BOND_PROFILE"
...
-bond.options: arp_interval=5,arp_ip_target=192.168.7.7,arp_validate=active,mode=balance-rr,use_carrier=0
+bond.options: mode=balance-rr,arp_interval=5,arp_ip_target=192.168.7.7,arp_validate=active,use_carrier=0
This doesn't affect keyfile, which sorts the hash keys themself (and
doesn't treat the "mode" special).
This however does affect ifcfg-rh writer how it writes the BONDING_OPTS
variable. I think this change is fine and preferable.
strcmp() is hard to understand visually. Especially when different patterns
are mixed, like:
if ( !strcmp (name, NM_SETTING_BOND_OPTION_MIIMON)
&& strcmp (value, "0") != 0) {
nm_utils_is_valid_iface_name() is a public API of libnm-core, let's use
our internal API.
$ sed -i 's/\<nm_utils_is_valid_iface_name\>/nm_utils_ifname_valid_kernel/g' $(git grep -l nm_utils_is_valid_iface_name)
In total, we register 447 property informations. Out of these,
326 are plain, GObject property based without special implementations.
The NMSettInfoProperty had all function pointers directly embedded,
currently this amounts to 5 function pointers and the "dbus_type" field.
That means, at runtime we have 326 times trivial implementations with
waste 326*6*8 bytes of NULL pointers. We can compact these by moving
them to a separate structure.
Before:
447 * 5 function pointers
447 * "dbus_type" pointer
= 2682 pointers
After:
447 * 1 pointers (for NMSettInfoProperty.property_type)
89 * 6 pointers (for the distinct NMSettInfoPropertType data)
= 981 pointers
So, in total this saves 13608 byes of runtime memory (on 64 bit arch).
The 89 NMSettInfoPropertType instances are the remaining distinct instances.
Note that every NMSettInfoProperty has a "property_type" pointer, but most of them are
shared. That is because the underlying type and the operations are the same.
Also nice is that the NMSettInfoPropertType are actually constant,
static fields and initialized very early.
This change also makes sense form a design point of view. Previously,
NMSettInfoProperty contained both per-property data (the "name") but
also the behavior. Now, the "behavioral" part is moved to a separate
structure (where it is also shared). That means, the parts that are
concerned with the type of the property (the behavior) are separate
from the actual data of the property.
If the mode is one of '802.3ad', 'tlb' or 'alb' and the connection has
both 'arp_interval' and 'arp_ip_target' options, during normalization
we remove 'arp_interval' because unsupported in the current mode. The
connection then becomes invalid because 'arp_ip_target' requires
'arp_interval'.
Since 'arp_interval' and 'arp_ip_target' are mutually dependent, the
latter should also be unsupported for those bonding modes.
https://bugzilla.redhat.com/show_bug.cgi?id=1718173
We no longer add these. If you use Emacs, configure it yourself.
Also, due to our "smart-tab" usage the editor anyway does a subpar
job handling our tabs. However, on the upside every user can choose
whatever tab-width he/she prefers. If "smart-tabs" are used properly
(like we do), every tab-width will work.
No manual changes, just ran commands:
F=($(git grep -l -e '-\*-'))
sed '1 { /\/\* *-\*- *[mM]ode.*\*\/$/d }' -i "${F[@]}"
sed '1,4 { /^\(#\|--\|dnl\) *-\*- [mM]ode/d }' -i "${F[@]}"
Check remaining lines with:
git grep -e '-\*-'
The ultimate purpose of this is to cleanup our files and eventually use
SPDX license identifiers. For that, first get rid of the boilerplate lines.
We have certain artificial properties that not only depend on one
property alone or that depend on a property in another(!) setting.
For that, we have synth_func.
Other than that, synth_func and get_func are really fundamentally
similar and should be merged. That is because the distinction whether a
property value is "synthetized" or just based on a plain property is
minor. It's better to have the general concept of "convert property to
GVariant" in one form only.
Note that compare_property() is by default implemented based
on get_func. Hence, if get_func and synth_func get merged,
compare_property() will also require access to the NMConnection.
Also it makes some sense: some properties are artificial and actually
stored in "another" setting of the connection. But still, the property
descriptor for the property is in this setting. The example is the
"bond.interface-name" which only exists on D-Bus. It's stored as
"connection.interface-name".
I don't really like to say "exists on D-Bus only". It's still a valid
property, despite in NMSetting it's stored somehow differently (or not
at all). So, this is also just a regular property for which we have a
property-info vtable.
Does it make sense to compare such properties? Maybe. But the point is that
compare_property() function needs sometimes access to the entire
connection. So add the argument.
Using strtol() correctly proves to be hard.
Usually, we want to also check that the end pointer is points to the end
of the string. Othewise, we silently accept trailing garbage.
Order the code in our common way. No other changes.
- ensure to include the main header first (directly after
"nm-default.h").
- reorder function definitions: get_property(), set_property(),
*_init(), *_new(), finalize(), *_class_init().
NMSetting's compare_property() has and had two callers:
nm_setting_compare() and nm_setting_diff().
compare_property() accepts a NMSettingCompareFlags argument, but
at the same time, both callers have another complex (and
inconsistent!) set of pre-checks for shortcuting the call of
compare_property(): should_compare_prop().
Merge should_compare_prop() into compare_property(). This way,
nm_setting_compare() and nm_setting_diff() has less additional
code, and are simpler to follow. Especially nm_setting_compare()
is now trivial. And nm_setting_diff() is still complicated, but
not related to the question how the property compares or whether
it should be compared at all.
If you want to know whether it should be compared, all you need to do
now is follow NMSettingClass.compare_property().
This changes function pointer NMSettingClass.compare_property(),
which is public API. However, no user can actually use this (and shall
not!), because _nm_setting_class_commit_full() etc. is private API. A
user outside of libnm-core cannot create his/her own subclasses of
NMSetting, and never could in the past. So, this API/ABI change doesn't
matter.
NMSetting internally already tracked a list of all proper GObject properties
and D-Bus-only properties.
Rework the tracking of the list, so that:
- instead of attaching the data to the GType of the setting via
g_type_set_qdata(), it is tracked in a static array indexed by
NMMetaSettingType. This allows to find the setting-data by simple
pointer arithmetic, instead of taking a look and iterating (like
g_type_set_qdata() does).
Note, that this is still thread safe, because the static table entry is
initialized in the class-init function with _nm_setting_class_commit().
And it only accessed by following a NMSettingClass instance, thus
the class constructor already ran (maybe not for all setting classes,
but for the particular one that we look up).
I think this makes initialization of the metadata simpler to
understand.
Previously, in a first phase each class would attach the metadata
to the GType as setting_property_overrides_quark(). Then during
nm_setting_class_ensure_properties() it would merge them and
set as setting_properties_quark(). Now, during the first phase,
we only incrementally build a properties_override GArray, which
we finally hand over during nm_setting_class_commit().
- sort the property infos by name and do binary search.
Also expose this meta data types as internal API in nm-setting-private.h.
While not accessed yet, it can prove beneficial, to have direct (internal)
access to these structures.
Also, rename NMSettingProperty to NMSettInfoProperty to use a distinct
naming scheme. We already have 40+ subclasses of NMSetting that are called
NMSetting*. Likewise, NMMetaSetting* is heavily used already. So, choose a
new, distinct name.
Previously, each (non abstract) NMSetting class had to register
its name and priority via _nm_register_setting().
Note, that libnm-core.la already links against "nm-meta-setting.c",
which also redundantly keeps track of the settings name and gtype
as well.
Re-use NMMetaSettingInfo also in libnm-core.la, to track this meta
data.
The goal is to get rid of private data structures that track
meta data about NMSetting classes. In this case, "registered_settings"
hash. Instead, we should have one place where all this meta data
is tracked. This was, it is also accessible as internal API,
which can be useful (for keyfile).
Note that NMSettingClass has some overlap with NMMetaSettingInfo.
One difference is, that NMMetaSettingInfo is const, while NMSettingClass
is only initialized during the class_init() method. Appart from that,
it's mostly a matter of taste, whether we attach meta data to
NMSettingClass, to NMMetaSettingInfo, or to a static-array indexed
by NMMetaSettingType.
Note, that previously, _nm_register_setting() was private API. That
means, no user could subclass a functioning NMSetting instance. The same
is still true: NMMetaSettingInfo is internal API and users cannot access
it to create their own NMSetting subclasses. But that is almost desired.
libnm is not designed, to be extensible via subclassing, nor is it
clear why that would be a useful thing to do. One day, we should remove
the NMSetting and NMSettingClass definitions from public headers. Their
only use is subclassing the types, which however does not work.
While libnm-core was linking already against nm-meta-setting.c,
nm_meta_setting_infos was unreferenced. So, this change increases
the binary size of libnm and NetworkManager (1032 bytes). Note however
that roughly the same information was previously allocated at runtime.
- Don't use @parent_class name. This local variable (and @object_class) is
the class instance up-cast to the pointer types of the parents. The point
here is not that it is the direct parent. The point is, that it's the
NMSettingClass type.
Also, it can only be used inconsistently, in face of NMSettingIP4Config,
who's parent type is NMSettingIPConfig. Clearly, inside
nm-setting-ip4-config.c we wouldn't want to use the "parent_class"
name. Consistently rename @parent_class to @setting_class.
- Also rename the pointer to the own class to @klass. "setting_class" is also the
wrong name for that, because the right name would be something like
"setting_6lowpan_class".
However, "klass" is preferred over the latter, because we commonly create new
GObject implementations by copying an existing one. Generic names like "klass"
and "self" inside a type implementation make that simpler.
- drop useless comments like
/* virtual functions */
/* Properties */
It's better to logically and visually structure the code, and avoid trival
remarks about that. They only end up being used inconsistently. If you
even need a stronger visual separator, then an 80 char /****/ line
should be preferred.
We commonly don't use the glib typedefs for char/short/int/long,
but their C types directly.
$ git grep '\<g\(char\|short\|int\|long\|float\|double\)\>' | wc -l
587
$ git grep '\<\(char\|short\|int\|long\|float\|double\)\>' | wc -l
21114
One could argue that using the glib typedefs is preferable in
public API (of our glib based libnm library) or where it clearly
is related to glib, like during
g_object_set (obj, PROPERTY, (gint) value, NULL);
However, that argument does not seem strong, because in practice we don't
follow that argument today, and seldomly use the glib typedefs.
Also, the style guide for this would be hard to formalize, because
"using them where clearly related to a glib" is a very loose suggestion.
Also note that glib typedefs will always just be typedefs of the
underlying C types. There is no danger of glib changing the meaning
of these typedefs (because that would be a major API break of glib).
A simple style guide is instead: don't use these typedefs.
No manual actions, I only ran the bash script:
FILES=($(git ls-files '*.[hc]'))
sed -i \
-e 's/\<g\(char\|short\|int\|long\|float\|double\)\>\( [^ ]\)/\1\2/g' \
-e 's/\<g\(char\|short\|int\|long\|float\|double\)\> /\1 /g' \
-e 's/\<g\(char\|short\|int\|long\|float\|double\)\>/\1/g' \
"${FILES[@]}"
constructor functions are ugly, because code is running before
main() starts. Instead, as the registration code for NMSetting types
is insid the GType constructor, we just need to ensure at the
right place, that the GType was created.
The right place here is _register_settings_ensure_inited(), because
that is called before we need the registration information.
NM_FLAGS_HAS() uses a static-assert that the second argument is a
single flag (power of two). With a single flag, NM_FLAGS_HAS(),
NM_FLAGS_ANY() and NM_FLAGS_ALL() are all identical.
The second argument must be a compile time constant, and if that is
not the case, one must not use NM_FLAGS_HAS().
Use NM_FLAGS_ANY() in these cases.
Make use of NMUtilsNamedValue in nm_utils_format_variant_attributes().
This avoids creating a GList and sorting it.
Also, reuse nm_utils_named_values_from_str_dict() in
nm_setting_bond_get_option().
We also do this for libnm and libnm-core, where it causes visible changes
in behavior. But if somebody would rely on the hashing implementation
for hash tables, it would be seriously flawed.
In an ideal world, we should not validate connections containing
options not valid for the current bond mode. However adding such
restriction now means that during an upgrade to the new NM version
some connections that were valid before become invalid, possibly
disrupting connectivity.
Instead, consider invalid options as a normalizable error and remove
them during normalization.
Converting the setting to a "canonical" form without invalid options
is important for the connection matching logic, where such invalid
options can cause false mismatches.
