KERNEL_MAKEOPTS will get expanded when it is used and not when it is
defined in the kernel.mk file now. This fixes problems finding dependent
kernel modules when it is used by a kernel module package.
Without this change the build of packages which depend on other out of
tree modules failed when they used KERNEL_MAKE because some symbols could
not be found. This happened because KERNEL_MAKE_FLAGS which contains a
"if $(__package_mk)" was evaluated where KERNEL_MAKEOPTS was defined
and not when the KERNEL_MAKE was used. For packages which included
kernel.mk before package.mk we saw this problem. One workaround
was to use the correct include order and the other one was to not
use KERNEL_MAKE_FLAGS, but copy its content.
Signed-off-by: Hauke Mehrtens <hmehrtens@maxlinear.com>
c3e31b6a9b and 5f8e587240 disable stack validation when the build
host is not running Linux, as the objtool kernel build tool required for
stack validation is not portable to other build host environments. This
was achieved by setting CONFIG_STACK_VALIDATION= in KERNEL_MAKEOPTS, and
by setting SKIP_STACK_VALIDATION=1 in the environment. KERNEL_MAKEOPTS
only has effect for the kernel build, not for external module builds,
but through kernel 5.14, SKIP_STACK_VALIDATION worked to disable this
feature too, so stack validation was disabled for external module builds
as well. Since kernel 0d989ac2c90b, the kernel build no longer considers
SKIP_STACK_VALIDATION, so the feature will be disabled for the kernel
build, but not for external module builds.
When building OpenWrt on a non-Linux build host targeting x86 (the only
target architecture for which OpenWrt enables the kernel
CONFIG_STACK_VALIDATION) and using kernel 5.15 (such as via
CONFIG_TESTING_KERNEL), this caused a build failure during any external
module build, such as kmod-button-hotplug. This manifested as build
errors such as:
make[4]: *** No rule to make target
'.../build_dir/target-x86_64_musl/linux-x86_64/button-hotplug/button-hotplug.o',
needed by
'.../build_dir/target-x86_64_musl/linux-x86_64/button-hotplug/button-hotplug.mod'.
Stop.
Although button-hotplug.c was present, the implicit rule to make
$(obj)/%.o from $(src)/%.c in the kernel's scripts/Makefile.build could
not be satisfied in this case, as it also depends on $(objtool_dep),
non-empty as a result of the failure to propagate disabling of stack
validation to external module builds, in a configuration where it is not
possible to build objtool.
KERNEL_MAKEOPTS is used for just the kernel build itself, while
KERNEL_MAKE_FLAGS is used for both the kernel build and for external
module builds. This restores the ability to build OpenWrt in such
configurations by moving the CONFIG_STACK_VALIDATION= make argument from
KERNEL_MAKEOPTS to KERNEL_MAKE_FLAGS where it is able to affect external
module builds properly.
Note that the kernel's objtool and related configuration have seen a
major overhaul since kernel 5.15, and may need more attention again
after 22922deae13f, in kernel 5.19.
Signed-off-by: Mark Mentovai <mark@mentovai.com>
Compiler option -no-plt will break kernel builds on some architectures
eg. (x86) Filter this option from the recently introduced handling of
KCFLAGS vs EXTRA_OPTIMISATION
Fixes: 1d42af720c ("kernel: use KCFLAGS for passing EXTRA_OPTIMIZATION flags")
Suggested-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
This uses kernel's generic variable and doesn't require patching it with
a custom Makefile change. It's expected *not* to change any behaviour.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
According to the GNU make manual, specifying library paths should be
done in LDFLAGS rather than LDLIBS. Replace KBUILD_HOSTLDLIBS with
KBUILD_HOSTLDFLAGS to pass the host lib directory.
Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
HOST_LOADLIBES was renamed to KBUILD_HOSTLDLIBS in kernel 4.19. As the
oldest kernel version we support is 5.10, cleanup HOST_LOADLIBES use.
Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
The license folder is a core part of OpenWrt and all GPL-2.0 licensed.
Use SPDX license tags to allow machines to check licenses.
Signed-off-by: Paul Spooren <mail@aparcar.org>
[rebase, keep some Copyright lines, sharpen commit message]
Signed-off-by: Adrian Schmutzler <freifunk@adrianschmutzler.de>
Commit 5d76065 moved the creation of the symvers directory to
include/kernel-build.mk. This is fine when building from scratch. But
when unpacking an SDK the directory doesn't exist and because the kernel
won't be built (again) this directory will not be created by the build
system, causing build failure if make tries to copy files into it.
This moves the creation of the symvers directory back into
include/kernel.mk so that the directory is created in any case.
Signed-off-by: Sebastian Kemper <sebastian_ml@gmx.net>
This replaces the previous (deprecated) method of collecting symvers data
in $(PKG_BUILD_DIR)/Module.symvers, which does not work on newer kernels
Signed-off-by: Felix Fietkau <nbd@nbd.name>
The symvers files of older kernel versions are incompatible with the ones
from 5.4, so changing the kernel version without running make clean was
causing build failures in kernel module packages.
Fix this by moving the directory, ensuring that symvers files get thrown
away with a kernel version change
Signed-off-by: Felix Fietkau <nbd@nbd.name>
In Linux kernel commit 8377bd2b9ee1 ("kbuild: Rename HOST_LOADLIBES to
KBUILD_HOSTLDLIBS") HOST_LOADLIBES was renamed to KBUILD_HOSTLDLIBS.
This patch adapts the OpenWrt kernel build to this new variable. Without
this change the kernel host tools would not link against the libraries
found in the staging directory.
Signed-off-by: Hauke Mehrtens <hauke.mehrtens@intel.com>
Some kmods (gpio-hotplug, wireguard) store the build path in the
compiled files and therefore make it harder to rebuild the official
binaries. As the same "iremap" function is used as for other binaries,
the change is compatible with gcc7 and 8.
Tested with both gcc7 and gcc8 resulting in build path independent
reproducible builds.
Signed-off-by: Paul Spooren <mail@aparcar.org>
The check for libelf in the kernel build is not enough, because the code that
uses libelf for stack validation is completely non-portable, as it tries to
include asm/types.h and relies on kernel types in user space.
Until this is fixed properly, the only solution is to disable this on any non
Linux build host
Signed-off-by: Felix Fietkau <nbd@nbd.name>
This changes the default PKG_BUILD_DIR to take BUILD_VARIANT into
account (if set), so that packages do not need to manually override
PKG_BUILD_DIR just to handle variants.
This also updates most base packages with variants to use the updated
default PKG_BUILD_DIR.
Signed-off-by: Jeffery To <jeffery.to@gmail.com>
Since we switched to 4.19, the kernel build checks for libelf to decide if
it should build tools for stack validation.
On macOS, this check fails during target/compile, but succeeds during package
build (because of the pkg-config path picking up target libraries).
Signed-off-by: Felix Fietkau <nbd@nbd.name>
e26ffb31df fixed only embedded modules
symbol collection. If we are building external modules, like broadcom-wl
or lantiq dsl stuff then modules which do EXPORT_SYMBOL have unresolved
paths in Module.symvers and external module which depend on other
external modules will have empty dependencies, leading to broken
module loading.
This was discussed on IRC with Jonas some time ago.
Fix this by handling both resolved and unresolved paths.
Fixes: e26ffb31df ("build: fix module symbol collection if build_dir is a symlink")
Signed-off-by: Roman Yeryomin <roman@advem.lv>
[jonas.gorski@gmail.com: add appropriate fixes tag]
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
When autoloading more than one modules per packages,
/etc/modules.d/$module depends on the file system ordering.
To test this: use disorderfs on the build_dir and build kmod-sched.
Signed-off-by: Alexander Couzens <lynxis@fe80.eu>
We need to use resolved file list as prerequisites for repacking kmod
.ipk files. Note that currently version_filter uses a Makefile macro
KERNEL_PATCHVER that should be available at ipk building time.
Reported-by: Rafał Miłecki <zajec5@gmail.com>
Signed-off-by: Yousong Zhou <yszhou4tech@gmail.com>
They are no longer stored in the "testing" subdirectory and are not
available as .tar.xz archives. If -rc is detected download it from the
git.kernel.org and use .tar.gz.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
The -rcX "testing" kernels are no longer hosted on
cdn.kernel.org file servers directly in a "testing"
directory. Therefore the logic that tested for "-rc"
can be removed.
Signed-off-by: Christian Lamparter <chunkeey@gmail.com>
kmod packages without FILES did not have an install step defined, leading
to no package being built. This affected netfilter/iptables packages, which
filter out builtin modules from FILES.
Not building a package that it is selected in .config is problematic, as
the generated empty package may be necessary to satisfy dependencies.
Signed-off-by: Matthias Schiffer <mschiffer@universe-factory.net>
Kernel 4.14 has the version number 4.14 and not 4.14.0. This was
different in some older Linux kernel versions, This change makes it
possible to use kernel 4.14 without any minor version.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
If PKG_BUILD_DIR contains symlinks, the generated Module.symvers will
contain the resolved paths, not the virtual path with the symlink name.
This breaks the filter for the module's own symbols, so to fix this
ensure we also grep for the resolved path.
Reported-by: Roman Yeryomin <roman@advem.lv>
Tested-by: Roman Yeryomin <roman@advem.lv>
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
Let the generic postinstall script invoke "kmodloader" when the just
installed package contains any /etc/module.d/ entries.
This allows us to skip the explicit "insert_module()" calls in the
package postinstall.
Due to the removed insert_module calls we do not need to assemble a
complete list of modules per package anymore, which allows for vast
simplification of the package generation code.
While we're at it, also support specifying default parameters for
modules using either the MODPARAM or MODPARAM.modulename variables
in KernelPackage.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
The creation was accidentally moved to the wrong hook, fix it by moving
it to the pre hook.
Fixes: e5e5c3f5fd ("build: ensure PKG_INFO_DIR exists before trying to use it")
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
PKG_INFO_DIR is only created at the finish step of the first package
build, but kernel modules now use it at the start. Doing a parallel build
could cause a kernel module to be the first package, breaking the build.
Fix this by ensuring the directory exists.
Fixes: 2e496876c6 ("kernel: collect module symvers for external modules")
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
Collect module symvers for all external modules to make them available
for modpost. This fixes dependencies for most external modules.
Example:
Before:
root@LEDE:/# modinfo mt76
module: /lib/modules/4.4.74/mt76.ko
license: Dual BSD/GPL
depends:
After:
root@LEDE:/# modinfo mt76
module: /lib/modules/4.4.74/mt76.ko
license: Dual BSD/GPL
depends: mac80211,cfg80211
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
During the initial configuration phases, we have not set-up the kernel
source directory, which would lead to such messages:
cat:
/local/users/fainelli/openwrt/trunk/build_dir/target-x86_64_musl/linux-uml/linux-4.9.58/include/config/kernel.release:
No such file or directory
Just silence it, since it does not create a functional problem.
Fixes: 8e0e0e7d8b ("include: Determine MODULES_DIR correctly for external/git kernels")
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
The kernel calls both ppc64 and ppc32 "powerpc", so we need to fixup
LINUX_KARCH when building with ARCH=powerpc64.
Signed-off-by: Florian Larysch <fl@n621.de>
This change currently causes some issues with loading out of tree kernel modules
so revert that commit for now.
Reverts commit 34c01e68b5. Fixes FS#919.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
Builtin modules are always present, and trying to load them will cause
modprobe to spew errors when installing the empty kmod packages.
Fix this by never generating any postinst module install instructions
for builtin modules.
Fixes#842.
Signed-off-by: Jonas Gorski <jonas.gorski@gmail.com>
When an external kernel tree is used the version should not get
modified by the LEDE build scripts. This was added by Florian some time
ago.
The commit 0aed054bec ("build: add KERNEL_MAKE and
KERNEL_MAKE_FLAGS variables and move to kernel.mk") breaks this feature
introduced in b6746a6ffb ("include: Do not alter KERNELRELEASE for
external/git kernels").
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
This allows packages to use kernel make options without the forced
-C $(LINUX_DIR). It also makes it more clear that it to be called from
kernel module packages directly.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
This commit contains the following changes
- Use local shell var where appliable
- The $(sort $$$$$$$$mods) call will have no expected effect
- Avoid EEXIST when creating symlinks in /etc/modules-boot.d/
- Avoid duplicate arguments for insert_modules() in postinst-pkg
Signed-off-by: Yousong Zhou <yszhou4tech@gmail.com>
When using external or git cloned kernels, any kind of modifications
will alter KERNELRELEASE. LEDE still tries to stage modules in
lib/modules/$(LINUX_UNAME_VERSION) and LINUX_UNAME_VERSION is based on
KERNEL_PATCHVER (indirectly) so this does not work, and we lose all
kinds of automatic modules loading.
To remedy that, just cat $(LINUX_DIR)/include/config/kernel.release
which is late enough the kernel has prepared this file, and is correctly
tracking changes done throughout the kernel.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Add a call to KernelPackage/$(1)/$(BOARD)/$(SUBTARGET) to the
KernelPackage macro. This allows to add kernel packages for x86/64,
without breaking x86. It's not possible to do this with BOARD, as
BOARD=x86 for x86_64.
Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
This is intended to be used for a wide array of package sanity checks.
The first check that is implemented is for the hash of downloaded files.
It checks:
- Missing hash
- Use of SHA256 instead of MD5
- dl/<file> hash not matching hash in makefile
- deprecated MD5SUM variable
The deprecated MD5SUM variable check is skipped for feeds/ until OpenWrt
is updated as well
Signed-off-by: Felix Fietkau <nbd@nbd.name>
This is to get reproducible builds of, eg, the kmod-sched ipkg.
Locale preferences can change build order, but the locale is already been
defined for the entire build process, so it doesn't need to be specified here.
Signed-off-by: bryan newbold <bnewbold@robocracy.org>
SVN-Revision: 48540
