So far, we were using a sort operation on the generated .config to
implement precedence of =y packages over =m, and =m over unset.
Unfortunately, this sort not only used for packages, but for all config
lines. This made it impossible to override settings from targets/generic
in a target config when the new setting was sorted before the generic
setting.
To fix this, track configurations by their keys, so we can properly
override config keys that were set before. Value-based precedence is
only preserved for package configuration.
The config() and try_config() calls always take key and value as
separate arguments now. Strings are quoted automatically; the values
true, nil and false map to y, m and unset for tristate options. config()
can take an optional third argument to override the error message to
display when the setting fails to apply.
All existing target configs generate the same .config with the old and the
new code. The new code is also a bit faster on targets with many devices.
The precedence of different package lists was broken since #1876,
disallowing removal of GLUON_FEATURES packages via GLUON_SITE_PACKAGES.
Including all package selections, both implicit defaults and explicit
handling in Gluon, the order of precedence is now the following:
1. OpenWrt defaults (including target-specific defaults)
2. Device-specific packages from OpenWrt
3. Generic default packages (from target/generic)
4. Target default packages (target/$(GLUON_TARGET))
5. Removal of opkg for tiny targets
6. Packages derived from GLUON_FEATURES + GLUON_FEATURES_$(class)
7. GLUON_SITE_PACKAGES
8. GLUON_SITE_PACKAGES_$(class)
9. Device-specific packages from target/$(GLUON_TARGET)
10. Device-specific packages from GLUON_$(device)_SITE_PACKAGES
This also contains various pieces of cleanup:
- No hardcoded order of device classes for target_config.lua arguments
anymore (in fact, the Makefile doesn't know anything about device
classes now)
- target_conifg_lib.lua only hardcodes the fallback class for x86, no
other occurences of specific class names
- Feature -> package list mapping is moved from Makefile to the Lua code
as well (still implemented in Shell though)
When adding device classes, targets without devices such as x86 were not
handled. As site and feature packages are included on such a per-device
decision, x86 images ended up without most packages.
Include a class setting for a target and include the class-packages
target-wide when this setting is configured.
Fixes 9c52365077 ("build: introduce device classes")
This commit allows to define a device-class flag in the target
definitions. This way, it is possible to distinguish between groups
of devices in the build-process in terms of package or feature
selection.
