OpenWrt packages

The OpenWrt system is maintained and distributed as a collection of packages.

Almost all pieces of software found in a typical OpenWrt firmware image are provided by such a package with a notable exception being the Linux kernel itself.

The term OpenWrt package may either refer to one of two things:

• an OpenWrt source package which essentially is a directory consisting of:
  • an OpenWrt package Makefile describing the acquisition, building and packaging procedures for a piece of software (required)
  • a supplemental directory with OpenWrt package patches which modify the acquired source code (optional)
  • other static files that go with the package, such as init script files, default configurations, scripts or other support files (optional)

• an OpenWrt binary package, which is a GNU tar compatible archive containing binary executable software artifacts and the accompanying package control files for installation on a running system, similar to the .deb or .rpm files used in other package managers

OpenWrt binary packages are almost exclusively produced from source packages by invoking either the OpenWrt buildroot or the OpenWrt SDK in order to translate the source package Makefile descriptions into executable binary artifacts tailored for a given target system.

Although it is possible to manually assemble binary packages by invoking tools such as tar and placing the appropriate control files in the correct directories, it is strongly discouraged to do so since such binary packages are usually not easily reproducible and verifiable.

Source packages are developed in multiple OpenWrt package feeds hosted in different locations and following different purposes. Each package feed is a collection of source package definitions residing within a publicly or privately reachable source code repository.

Source packages describe how a piece of software has to be downloaded, patched, compiled and packaged in order to form a binary software artifact suitable for use on a running target system. They also describe relations to other source packages required either at build time or at run time.

Each source package should have a globally unique name closely resembling the name of the software described by it. OpenWrt often follows the lead of other distributions when deciding about the naming of packages and sticks to the same naming conventions in many cases.

A source package is a subdirectory within its corresponding package feed containing at least one Openwrt Makefile and optionally src, files or patches directories.

An OpenWrt source package Makefile contains a series of header variable assignments, action recipes and one or multiple OpenWrt specific signature footer lines identifying it as OpenWrt specific package Makefile.

See Creating packages for details on Makefile contents.

Static files accompanying a source package, such as OpenWrt specific init scripts or configuration files, must be placed inside a directory called files, residing within the same subdirectory as the Makefile. There are no strict rules on how such static files are to be named and organized within the files directory but by convention, the extension .conf is used for OpenWrt UCI configration files and the extension .init is used to denote OpenWrt specific init scripts.

The actual placement and naming of the resources within the files directory on the target system is controlled by the source package Makefile and unrelated to the structure and naming within the files directory.

The patches directory must be placed in the same parent directory as the Makefile file and may only contain patch files used to modify the source code being packaged.

Patch files must be in unified diff format and carry the extension .patch. The file names must also carry a numerical prefix to denote the order in which the patch files must be applied. Patch file names should be concise and avoid characters other than ASCII alphanumerics and hyphens.

Suitable patch file names could look like:

• 000-patch-makefile.patch
• 010-backport-frobnicate-crash-fix.patch
• 999-add-local-hack-for-openwrt-compatibility.patch

It is recommended to use Quilt to manage source package patch collections.

Some packages do not actually fetch their program code from an external source but bundle the code to be compiled and packages directly within the package feed. This is usually done for packages which are specific to OpenWrt and do not exist outside of their respective package feed.

Sometimes the src directory may also be used to supply additional code to the compilation process, in addition to the program code fetched from external sources.

If present, the OpenWrt build system will automatically copy the contents of the src directory verbatim to the compilation scratch directory (build directory) of the package, retaining the structure and naming of the files.

Many OpenWrt supported devices still have only a few megabytes of flash and RAM available which makes it important to shrink the packages as much as possible. Opt for the lowest common denominator whenever possible.

Some general considerations when packaging a new piece of software are:

• Do not ship man pages or documentation, a typical installation lacks both the infrastructure and the space to view and store man page databases
• Minimize external dependencies - try to avoid optional external dependencies whenever possible. An extreme example is ICU which weighs around 12MB and is an optional dependency for Unicode multi language support in various packages
• Modularize packages - if the software you're packaging supports and uses plugins then put those plugins into separate binary package declarations instead of lumping them all together along with the main program. This way you can externalize dependencies and move them into the plugin packages instead of having them in the main component, which makes the package usable on a wider range of targets because users can omit parts with large dependencies.
• Try to rely on standard facilities - instead of requiring extra programs to implement tasks like user context switching, use the procd facilities to run a service as a different user.

Often it is tempting to add various menuconfig configuration options to allow the customization of the package features by the users compiling their own variant of OpenWrt but it should be kept in mind that large parts of the userbase will use the package solely by installing binary archives from the OpenWrt repositories.

Binary packages in the official OpenWrt repositories are always built with the default settings of a package so a maintainer should ensure that the default feature selection represents a fair balance between resource requirements and most common user needs.

Historically, packages for OpenWrt used to contain a copyright notice at the top of the Makefile, stating something like:

# Copyright (C) 2007-2010 OpenWrt.org
This is free software, licensed under the GNU General Public License v2.
See /LICENSE for more information.

Since contributors likely do not have a formal contract with OpenWrt to develop packages, they cannot disclaim their own copyrights and assign them to the project.

When adding new packages, please don't simply copy the statement from another package but add either your own in the form:

# Copyright (C) 2016 Joe Random <joe@example.org>

or omit it entirely.

There are a number of Makefile variables influencing the visible version of the resulting packages. When packaging upstream release tarballs, the PKG_VERSION variable should be set to the version of the upstream software being packaged. For example, if the openssl package compiles the released openssl-1.0.2q.tar.gz archive, then PKG_VERSION variable should be set to the value 1.0.2q.

When there are no upstream release tarballs available or when software is packaged straight from a source code repository, the PKG_SOURCE_DATE and PKG_SOURCE_VERSION variables should be used instead. The PKG_SOURCE_DATE value must correspond to the modification date in the format YYYY-MM-DD of the source repository revision being packaged and PKG_SOURCE_VERSION must be set to the corresponding revision identifier of the repository, e.g. the commit hash for Git or the revision number for SVN repositories. For example, if the ubus package clones from Git revision https://git.openwrt.org/?p=project/ubus.git;a=commitdiff;h=221ce7e7ff1bd1a0c9995fa9d32f58e865f7207f, then its Makefile should specify PKG_SOURCE_DATE:=2018-10-06 and PKG_SOURCE_VERSION:=221ce7e7ff1bd1a0c9995fa9d32f58e865f7207f.

The build system will combine these variables into a common version identifier and truncate the revision identifier if needed. Given the values in the example, the resulting version identifier will be 2018-10-06-221ce7e7. This is done to make repository revision identifiers comparable to each other since SCM systems like Git or Mercurial use SHA hashes to identify revisions which are no monotonically increasing numerical values.

Source packages must specify a PKG_RELEASE value identifying the revision of the source package. In contrast to the PKG_VERSION, PKG_SOURCE_DATE and PKG_SOURCE_VERSION variables which are identifying the upstream version of the program code being packaged, the PKG_RELEASE variable denotes the revision of the package itself.

The package revision should start with the value 1 and must be increased whenever modifications are made to the package which might cause changes to the executables or other files contained within the resulting binary packages. When the package is updated to a newer PKG_VERSION or PKG_SOURCE_VERSION, the PKG_RELEASE must be reset back to 1.

Some examples for dealing with the PKG_RELEASE are:

• Fixed a typo in the maintainer's mail address → PKG_RELEASE stays unchanged
• Added a --disable-acl to the configure arguments → PKG_RELEASE is incremented
• Updated libfoo from version 0.2.1 to 0.2.2 → PKG_RELEASE is reset to 1 and PKG_VERSION set to 0.2.2

When declaring the source download method in the Makefile, direct tarball downloads via HTTP or HTTPS are the preferred way to acquire package sources, Git or other SCM clones should be avoided, mainly to keep the locally cached source downloads reproducible.

If direct Git cloning is required (for example because there is no release tarballs available upstream) then Git via HTTPS is preferred over Git via HTTP is preferred over Git via its native protocol. Many OpenWrt users are behind corporate firewalls which disallow Git native traffic (TCP 9418).

The use of mirror sites for tarball download locations is encouraged and helps to reduce the traffic load on upstream project sites. When choosing mirrors for a package, please try to ensure that the mirror is:

• officially endorsed by the upstream project (E.G. mentioned on their download page).
• well reachable by people from a wide range of different locations.
• using proper SSL certificates when using HTTPS.
• hosting the most current version of the software in question.

Multiple mirrors can be specified in a package Makefile by assigning a white-space separated list of URLs to the PKG_SOURCE_URL variable. It is a good convention to assign the upstream project site itself to the end of the mirror list. This provides a canonical fallback location in case a new version has not yet propagated to all mirrors and conveys the original download location to casual readers.

Try to limit the amount of mirror sites to 3 to 5 different locations, including the main download site.

The build recipes in a source package should adhere to the OpenWrt defaults as much as possible. This ensures that source package declarations remain compact and free of copy-pasted boilerplate code.

By default, the build system uses a set of standard ./configure and make invocations to build packages in a refined manner. Most of these steps can be influenced through a number of variables to alter the way the actual commands are executed.

Please refer to package-defaults.mk to learn how the default recipes are implemented.

Whenever possible, try to avoid redefining the default macros but use the provided variables to encode functional differences.

Example for a bad redefinition:

define Build/Compile 
        (cd $(PKG_BUILD_DIR)/nonstandard/dir/; make)
endef

Example for achieving the same using variable overrides:

MAKE_PATH := nonstandard/dir/

Likewise, do not attempt to override Build/Configure but use CONFIGURE_ARGS to pass switches like CONFIGURE_ARGS += --enable-acl or CONFIGURE_ARGS += --without-systemd and CONFIGURE_VARS to pass environment variables to the configuration script, like CONFIGURE_VARS += ac_cv_func_snprintf=yes.

In some cases it is possible to arrange things before e.g. the ./configure script is invoked in order to touch files, remove things or echo values into stampfiles. In such cases, it is permissible to redefine the recipe in order to achieve the desired result. Use the default implementations of the macros to call the original behaviour after the custom work is done. Refer to the examples below for some common use cases.

define Build/Prepare
	echo "1.2.3" > $(PKG_BUILD_DIR)/version.txt
	$(call Build/Prepare/Default)
endef

define Build/Prepare
	$(call Build/Prepare/Default)
	rm -f $(PKG_BUILD_DIR)/m4/libtool.m4
	cp $(PKG_BUILD_DIR)/make/Makefile.linux $(PKG_BUILD_DIR)/Makefile
endef

define Build/Configure
	touch $(PKG_BUILD_DIR)/ChangeLog
	$(call Build/Configure/Default)
endef

define Build/Compile
	$(call Build/Compile/Default)
	cp $(PKG_BUILD_DIR)/src/libfoo.so.1.2 $(PKG_BUILD_DIR)/src/libfoo.so
endef

Many open source projects rely on GNU autoconf and automake as their build system which may lead to a number of problems in a cross compilation setting.

Usual problems revolve around:

• configure scripts attempting to call programs to test certain features which might fail if the called program has been built for another architecture
• Pregenerated configure scripts embedding faulty and possibly outdated versions of libtool causing runtime problems on the target system
• Macros in configure scripts probing host system details to configure the package for the target, like calling uname to figure out the kernel version or endianess
• Projects shipping convenience scripts like autogen.sh which make certain assumptions about the host system or try to call the improper version of utilities like autopoint or autoconf, leading to macro errors and version mismatches when executing the generated configure scripts and Makefiles

Due to the complex nature of the GNU autoconf/automake system there is no single set of solutions to a given problem but rather a general list of guidelines and best practices to adhere to.

• Never patch the generated / shipped configure script but fix the underlying configure.ac or configure.in files and rely on the PKG_FIXUP:=autoreconf facility to regenerate the config script. This also has the nice side-effect of updating the embedded libtool version and using a cross-compile-safe set of standard macros, replacing unsafe ones in many cases.
• Make ./configure invocations as explicit as possible by forcibly disabling or enabling any feature which depends on the presence of an external library, e.g. --disable-acl to build a given package without libacl support on both systems having libacl in their staging directory and systems not providing this library. Failure to do so will result in errors like Package example is missing dependencies for the following libraries: libfoo.so.1 on systems that happen to build libfoo before building example.
• Pre-seed configure tests that cannot be reliably determined in a cross-compile setting. Properly written autoconf test macros can be overridden by cache-variables in the form ac_cv_somename=value - use this facility to skip tests which would otherwise fail or result in host-system specific values. For example the libpcap package passes ac_cv_linux_vers=$(LINUX_VERSION) to prevent ./configure from calling the host systems uname in order to figure out the kernel version. The names of the involved cache variables can be found in the config.log file within the package build directory or by inspecting the generated shell code of the configure program. Use the CONFIGURE_VARS variable to pass the cache variables down to the actual ./configure invocation
• Never trust shipped autogen.sh and similar scripts, rather use PKG_FIXUP:=autoreconf to (re)generate the configure script and automake templates and encode additionally needed steps in the appropriate build recipes.

A source package may depend on a number of other packages, either to satisfy compilation requirements or to enforce the presence of specific functionality, such as shared libraries or support programs at runtime on the target system.

There are two kinds of dependencies; build dependencies, specified by the PKG_BUILD_DEPENDS Makefile variable and runtime dependencies, declared in the DEPENDS variable of the corresponding define Package/... Makefile sections.

Build dependencies are resolved at package compilation time and instruct the build system to download, patch and compile each mentioned dependency before the source package itself is compiled. This is required when the compilation process of a package depends on resources such as header files from another package. Build dependencies are not transformed into runtime dependencies and should only be used when the resources of the packages being depended upon are solely required at compilation time. This usually is the case for header-only libraries such as the C++ Boost project or static .a library archives that result in no dynamic runtime requirements.

Runtime dependencies, on the other hand, specify the relation of binary packages, instructing the package manager to fetch and install the listed dependencies before installing the binary package itself. A runtime dependency automatically implies a build dependency, so if a DEPENDS variable within a define Package/... section of a given source package specifies the name of a define Package/... section of another source package, the build system will take care of compiling the other package before compiling the source package specifiying the runtime dependency.

Package dependencies, regardless of whether they're build-time or runtime ones, should only require packages within the same package feed or provided by the base feed located within the main OpenWrt package/ directory.

Dependencies among packages in different, non-base feeds are strongly discouraged as it is not guaranteed that each build system has access to all feeds.

Packages providing shared libraries require additional care to ensure that software depending on these libraries remains functional and is not accidentally broken by incompatible updates, changed APIs, removed functionality and so on.

While the package dependency mechanisms will ensure that the build system compiles library packages before the program packages requiring them, they do not guarantee that such programs are getting rebuilt when the library package itself is updated.

Also, in the case of binary package repositories, installing a newer, incompatible version of library packages would break installed programs relying on this library unless an additional version constraint is applied to the dependency.

The OpenWrt build system introduced the concept of an ABI_VERSION to address the issue of program dependencies on specific versions of a shared library, requiring exactly the ABI the program was initially compiled and linked against. The ABI_VERSION value is supposed to reflect the SONAME of the library being packaged.

Most upstream libraries contain an ELF SONAME attribute denoting the canonical name of the library including a version suffix specifying the version of the exposed ABI. Changes breaking the exposed ABI usually result in a change to the SONAME.

When a program is linked against such a library, the linker will resolve the SONAME of the requested shared object and put it into the DT_NEEDED section of the resulting program executable. Upon starting the program, the dynamic linker on the target system will consult the DT_NEEDED section to find the required libraries within the standard library search path.

Setting an ABI_VERSION variable on a library package definition will cause the build system to track the value of this variable and trigger recompilations in all packages depending on this library package whenever the value is incremented. This is useful to force re-linking of all programs after a library has been changed to an incompatible version.

The ABI_VERSION value is also appended to the binary package name and all dependencies mentioning the binary library package will be automatically expanded to contain the ABI_VERSION suffix. If for example a library package libfoo specifies ABI_VERSION:=1.0, the resulting binary package will be called libfoo1.0 and when a package bar specifies DEPENDS:=+libfoo, the resulting runtime dependency will be Depends: libfoo1.0.

This ensures that incompatible updates to the libfoo library, denoted by an ABI_VERSION increase, will cause programs linked against it from then on to have a different runtime dependency, allowing the OpenWrt package manager to notice the change.

Example: when libfoo is updated to a new, incompatible version and its SONAME property changed from libfoo.so.1.0 to libfoo.so.1.1, then ABI_VERSION should be increased from 1.0 to 1.1, causing the resulting libfoo binary package to be called libfoo1.1. Source packages linking libfoo from then on, will have runtime dependencies on libfoo1.1.

When a shared library is packaged, the ABI_VERSION variable of the corresponding define Package/lib... section should be set to the SONAME of the .so library file contained within the binary package. The SONAME usually reflects the library's internal ABI version and is incremented whenever incompatible changes to the public APIs are made within the library, E.G. when changing a function call signature or when removing exported symbols.

The public ABI tracker is useful to decide whether an ABI_VERSION change is required when updating an existing library package to a newer upstream version.

Some upstream library projects do not use a SONAME at all or do not properly version their libraries, in such cases, the ABI_VERSION must be set to a value in the form YYYYMMDD, reflecting the source code change date of the last incompatible change being made.

In order to allow multiple versions of binary library packages to coexist on the same system, each library package should only contain shared object files specific to the SONAME version of the library getting packaged.

A typical upstream library libbar with version 1.2.3 and SONAME of libbar.so.1 will usually provide these files after compilation:

libbar.so       -> libbar.so.1.2.3 (symlink)
libbar.so.1     -> libbar.so.1.2.3 (symlink)
libbar.so.1.2.3      (shared library object)

The binary libbar1 package should only contain libbar.so.1 and libbar.so.1.2.3 as the common libbar.so symlink would clash with a libbar2 package providing version 2.0.0 of libbar.

Versionless symlinks are usually not needed for libraries using the SONAME attribute and are only used during the linking phase when compiling programs depending on the library.

$(INSTALL_DATA) and $(INSTALL_BIN) will currently copy the file contents instead of the symlink itself, so prefer $(CP) when copying the library symlinks. Consider the example above, if you run

$(INSTALL_BIN) $(PKG_INSTALL_DIR)/usr/lib/libbar.so.* $(1)/usr/lib/

it will result in two copies of the library in regular files:

libbar.so.1               (regular file)
libbar.so.1.2.3           (regular file)

Instead, use

$(CP) $(PKG_INSTALL_DIR)/usr/lib/libbar.so.* $(1)/usr/lib/

and you'll get the intended result:

libbar.so.1     -> libbar.so.1.2.3 (symlink)
libbar.so.1.2.3                    (regular file)

While there has been a proposal to change $(INSTALL_BIN) behavior, $(CP) will continue to work.

Source packages defining binary packages that ship shared libraries should declare a Build/InstallDev recipe that copies all resources required to discover and link the shared libraries into the staging directory.

A typical InstallDev recipe usually copies all library symlinks (including the unversioned ones), header files and, in case they're provided, pkg-config (*.pc) files.