NetworkManager/CONTRIBUTING.md

18 KiB

Guidelines for Contributing

Community

Check out website https://networkmanager.dev and our GNOME page.

The release tarballs can be found at download.gnome.org.

Our mailing list is networkmanager@lists.freedesktop.org (archive, old-archive).

Find us on IRC channel #nm on Libera.Chat.

Report issues and send patches via gitlab.freedesktop.org or our mailing list.

Documentation

Find the documentation on our website.

NetworkManager is partly licensed under terms of GNU Lesser General Public License version 2 or later (LGPL-2.1-or-later). That is for example the case for libnm. For historical reasons, the daemon itself is licensed under terms of GNU General Public License, version 2 or later (GPL-2.0-or-later). See the SPDX license comment in the source files.

Note that all new contributions to NetworkManager MUST be made under terms of LGPL-2.1-or-later, that is also the case for files that are currently licensed GPL-2.0-or-later. The reason is that we might one day use the code under terms of LGPL-2.1-or-later and all new contributions already must already agree to that. For more details see RELICENSE.md.

Do not use "Signed-off-by:" lines in commits for NetworkManager. It has no meaning.

Coding Style

clang-format

The formatting is automated using clang-format. Run ./contrib/scripts/nm-code-format.sh -i ([1]) to reformat the code or run clang-format directly.

As the generated format depends on the version of clang-format, you need to use the correct clang-format version. That is basically the version that our gitlab-ci pipeline uses for the "check-tree" test. This is the version from a recent Fedora installation.

You may also run ./contrib/scripts/nm-code-format-container.sh which uses a Fedora container with podman and the correct version of clang-format.

You are welcome to not bother and open a merge request with wrong formatting, but note that we then will automatically adjust your contribution before merging.

The automatic reformatting was done by commit 328fb90f3e. Use --ignore-rev option or --ignore-revs-file .git-blame-ignore-revs to ignore the reformatting commit with git-blame:

$ git config --add 'blame.ignoreRevsFile' '.git-blame-ignore-revs'

You may integrate clang-formatter in your editor (for vim).

Style

As we use clang-format, our style is in parts determined by the tool. Run the tool to format the code. See the earlier point.

The formatting tool cannot cover all questions. The most important rule is to mimic the existing code and imitate the surrounding style.

In general, we require to build without compiler warnings, for the warnings that we enable. Our language is C11 with some GCC-isms (like typeof(), expression statements, cleanup attribute). In practice, we support various versions of GCC and clang. The supported C "dialect", compilers and libc are those that we can practically build and test in our CI. We don't target a theoretical, pure C11/POSIX standard or a libc/compiler that we cannot test. Patches for making NetworkManager more portable are welcome, if there is a practical use and checked by CI. Glibc and musl libc are supported.

We follow a mixture of glib's and systemd's style, which already have extensive guidelines. Following there are a few noteworthy points.

  • Use cleanup functions (gs_free, gs_*, nm_auto*) to let a stack variable own a resource instead of explicit free. Combine them with g_steal_pointer() to transfer ownership and with clear functions (g_clear_object(), nm_clear_g_free(), nm_clear*()) to destroy the resource early.

  • Use GSource instances instead of the source IDs from g_idle_add(), g_timeout_add(), etc. Possibly use nm_g_idle_add_source(), nm_g_timeout_add_source(), etc. and combine with nm_clear_g_source_inst().

  • Don't use GDBusProxy or GDBusObjectManager. Use plain GDBusConnection.

  • Names in our header files should always have an "nm" prefix (like "nm_", "NM_", "nm", "nmp"). Names in source files usually should not have an "nm" prefix.

  • Indent with spaces. (no tabs).

  • C-style comments

    • GOOD: f(x); /* comment */
    • BAD: f(x); // comment
  • Keep assignments in the variable declaration area pretty short.

    • GOOD: MyObject *object;
    • BAD: MyObject *object = complex_and_long_init_function(arg1, arg2, arg3);
  • Declare each variable on a separate line:

    • BAD: int i, j;
  • Constants are CAPS_WITH_UNDERSCORES and use the preprocessor.

    • GOOD: #define MY_CONSTANT 42
    • BAD: static const unsigned myConstant = 42;
  • Always use curly braces for blocks that span multiple lines. For single lines the braces may be omitted, but are not prohibited.

Checkpatch

We have a checkpatch.pl script, which is also run in our gitlab-ci. Review the warnings, but as these are just heuristics, there might be valid reasons to reject them. There is also a git hook which you can call from .git/hooks/post-commit.

Building from Source

First see that you have the required build dependencies. For Fedora/RHEL/Centos, you can look at this script and here is a script for Debian/Ubuntu.

Both meson and autotools are supported. You may choose whatever you prefer. For autotools the common steps are

./autogen.sh $CONFIGURE_OPTIONS
make -j 8
# optional: sudo make install

and for meson it's

meson build $CONFIGURE_OPTIONS
ninja -C build
# optional: sudo meson install -C build

Beware to set the correct $CONFIGURE_OPTIONS. In particular, you may not want the default installation prefix and not overwrite files in /usr.

Fedora

For Fedora/RHEL/CentOS, you can build an RPM from upstream sources with

   ./contrib/fedora/rpm/build_clean.sh -r

Pass --help to build_clean.sh for options.

You may also use the Copr project maintained by the upstream maintainers. There you find builds of latest main and stable branches.

Unit Tests

We have plenty of unit tests. Run them with make check or meson test -C build.

Note that some files in the source tree are both generated and commited to git. That means, certain changes to the code also affect these generated files. The unit test fail in that case, to indicate that the generated files no longer match what is commited to git. You can also automatically regenerate the files by running NM_TEST_REGENERATE=1 make check. Note that test-client requires working translation. See the comment for how to configure it.

Code Structure

./contrib- Contains a lot of required package, configuration for different platform and environment, build NM from source tree.

./data- Contains some configurations and rules.

./docs- Contains the generated documentation for libnm and for the D-Bus API.

./examples- Some code examples for basic networking operations and status checking.

./introspection- XML docs describing various D-Bus interface and their properties.

./m4- Contains M4 macros source files for autoconf.

./man- NM manual files.

./po- contains text-based portable object file. These .PO files are referenced by GNU gettext as a property file and these files are human readable used for translating purpose.

./src- source code for libnm, nmcli, nm-cloud-setup, nmtui…

./tools- tools for generating the intermediate files or merging the file.

Cscope/ctags

NetworkManager's source code is large. It may be a good idea to use tools like cscope/ctags to index the source code and navigate it. These tools can integrate with editors like Vim and Emacs. See:

For cscope, you can also set $SOURCEDIRS to include other source trees and navigate those sources. For example,

export SOURCEDIRS=/path/to/glib:/path/to/libndp
cscope -b -q -R -ssrc

Miscellaneous

Assertions in NetworkManager code

There are different kind of assertions. Use the one that is appropriate.

  1. g_return_*() from glib. This is usually enabled in release builds and can be disabled with G_DISABLE_CHECKS define. This uses g_log() with G_LOG_LEVEL_CRITICAL level (which allows the program to continue, unless G_DEBUG=fatal-criticals or G_DEBUG=fatal-warnings is set). As such, this is usually the preferred way for assertions that are supposed to be enabled by default.

    Optimally, after a g_return_*() failure the program can still continue. This is also the reason why g_return_*() is preferable over g_assert(). For example, that is often not the case for functions that return a GError, because g_return_*() will return failure without setting the error output. That often leads to a crash immediately after, because the caller requires the GError to be set. Make a reasonable effort so that an assertion failure may allow the process to proceed. But don't put too much effort in it. After all, it's an assertion failure that is not supposed to happen either way.

  2. nm_assert() from NetworkManager. This is disabled by default in release builds, but enabled if you build --with-more-assertions. See the WITH_MORE_ASSERTS define. This is preferred for assertions that are expensive to check or nor necessary to check frequently. It's also for conditions that can easily be verified to be true and where future refactoring is unlikely to break the invariant. Use such asserts deliberately and assume they are removed from production builds.

  3. g_assert() from glib. This is used in unit tests and commonly enabled in release builds. It can be disabled with G_DISABLE_ASSERT define. Since such an assertion failure results in a hard crash, you should almost always prefer g_return_*() over g_assert() (except in unit tests).

  4. assert() from C89's <assert.h>. It is usually enabled in release builds and can be disabled with NDEBUG define. Don't use it in NetworkManager, it's basically like g_assert().

  5. g_log() from glib. These are always compiled in, depending on the logging level they act as assertions too. G_LOG_LEVEL_ERROR messages abort the program, G_LOG_LEVEL_CRITICAL log a critical warning (like g_return_*(), see G_DEBUG=fatal-criticals) and G_LOG_LEVEL_WARNING logs a warning (see G_DEBUG=fatal-warnings). G_LOG_LEVEL_DEBUG level is usually not printed, unless G_MESSAGES_DEBUG environment variable enables it.

    In general, avoid using g_log() in NetworkManager. We have nm-logging instead which logs to syslog or systemd-journald. From a library like libnm it might make sense to log warnings (if something is really wrong) or debug messages. But better don't. If it's important, find a way to report the condition via the API to the caller. If it's not important, keep silent. In particular, don't use levels G_LOG_LEVEL_CRITICAL and G_LOG_LEVEL_WARNING because we treat them as assertions and we want to run all out tests with G_DEBUG=fatal-warnings.

  6. g_warn_if_*() from glib. These are always compiled in and log a G_LOG_LEVEL_WARNING warning. Don't use this.

  7. G_TYPE_CHECK_INSTANCE_CAST() from glib. Unless building with WITH_MORE_ASSERTS, we set G_DISABLE_CAST_CHECKS. This means, cast macros like NM_DEVICE(ptr) translate to plain C pointer casts. Use such cast macros deliberately, in production code they are cheap, with more asserts enabled they check that the pointer type is suitable.

Of course, every assertion failure is a bug, and calling it must have no side effects.

Theoretically, you are welcome to set G_DISABLE_CHECKS, G_DISABLE_ASSERT and NDEBUG in production builds. In practice, nobody tests such a configuration, so beware.

For testing, you also want to run NetworkManager with environment variable G_DEBUG=fatal-warnings to crash upon G_LOG_LEVEL_CRITICAL and G_LOG_LEVEL_WARNING g_log() message. NetworkManager won't use these levels for regular logging but for assertions.

Header Includes

Almost all C source header should include a certain set of default headers. That comes from the fact, that (almost) all source should include autotools' "config.h" as first.

That means, (almost) all our C sources should start with

/* SPDX-License-Identifier: LGPL-2.1-or-later */

#include "$BASE/nm-default$EXTRA.h"

that is, the first header is one of the several "*/nm-default*.h" headers. This header ensure that certain headers like libnm-std-aux/nm-std-aux.h and basics like nm_assert() and nm_auto_g_free are available everywhere.

The second include is the header that belongs to the C source file. This is so that header files are self-contained (aside what default dependencies that they get and everybody can rely on).

The next includes are system headers with <>. Exceptions are headers like "libnm-std-aux/nm-linux-compat.h" and "nm-compat-headers/*" which are our small wrappers around system headers. These are also to be included together with system headers.

Finally, all other headers from our source tree. Note that all build targets have -I. -I./src/ in their build arguments. So to include a header like src/libnm-glib-aux/nm-random-utils.h you'd do #include "libnm-glib-aux/nm-random-utils.h".

Note that there are exceptions. For example, src/libnm-std-aux/nm-linux-compat.h](src/libnm-std-aux/nm-linux-compat.h) may need to be included before system headers as it is supposed to include headers from src/linux-headers](src/linux-headers).

See an example here.

GObject Properties

We use GObjects and GObject Properties in various cases. For example:

  1. In public API in libnm they are used and useful for providing a standard GObject API. One advantage of GObject properties is that they work well with introspection and bindings.

  2. NMSetting properties commonly are GObject properties. While we provide C getters, they commonly don't have a setter. That is, settings can often only set via g_object_set().

  3. Our D-Bus API uses glue code. For the daemon, this is nm-dbus-manager.[ch] and nm-dbus-object.[ch]. For libnm's NMClient, this is nm-object.c. These bindings rely on GObject properties.

  4. Sometimes it is convenient to use the functionality that GObject properties provide. In particular, notify:: property changed signals or the ability to freeze/thaw the signals.

  5. Immutable objects are great, so there is a desire to have G_PARAM_CONSTRUCT_ONLY properties. In that case, avoid adding a getter too, the property only needs to be writable and you should access it via the C wrapper.

In general, use GObject properties sparsely and avoid them (unless you need them for one of the reasons above).

Almost always add a #define for the property name, and use for example g_signal_connect(obj, "notify::"NM_TARGET_SOME_PROPERTY", ...). The goal is to be able to search the use of all properties.

Almost always add C getter and setters and prefer them over g_object_get() and g_object_set(). This also stresses the point that you usually wouldn't use a GObject property aside the reasons above.

When adding a GObject properties, do it for only one of the reasons above. For example, the property NM_MANAGER_DEVICES in the daemon is added to bind the property to D-Bus. Don't use that property otherwise and don't register a notify::NM_MANAGER_DEVICES for your own purpose. The reason is that GObject properties are harder to understand and they should be used sparsely and for one specific reason.

Git Notes (refs/notes/bugs)

We use special tags in commit messages like "Fixes", "cherry picked from" and "Ignore-Backport". The find-backports script uses these to find patches that should be backported to older branches. Sometimes we don't know a-priory to mark a commit with these tags so we can instead use the bugs notes.

The git notes reference is called "refs/notes/bugs".

So configure:

$ git config --add 'remote.origin.fetch' 'refs/notes/bugs:refs/notes/bugs'
$ git config --add 'notes.displayref' 'refs/notes/bugs'

For example, set notes with

$ git notes --ref refs/notes/bugs add -m "(cherry picked from commit $COMMIT_SHA)" HEAD

You should see the notes in git-log output as well.

To resync our local notes use:

$ git fetch origin refs/notes/bugs:refs/notes/bugs -f

Testing NetworkManager with nm-in-container script.

See the readme for details.