[PATCH] build: fix regression for kernels < 5.10

[Sebastian Kemper]

On Tue, Apr 13, 2021 at 03:09:24PM +0200, Felix Fietkau wrote:
>
> On 2021-04-13 14:22, Sebastian Kemper wrote:
> > This fixes a regression introduced with commit
> > 5ed1e5140a80558ab47fd70410ae3242bed5becf ("build: build kernel image
> > before building modules/packages").
> >
> > Before this commit the make target would always include "modules",
> > resulting in a MODPOST and a complete Module.symvers file. Since this
> > commit a MODPOST of the kernel modules is not guaranteed for kernels <
> > 5.10. This results in some broken SDKs in which external packages that
> > depend on exported symbols from kernel modules fail to compile.
> Why is it not enough to do this in the CompileModules step?

Because Module.symvers gets regenerated during CompileImage as well.

>
> > Adding "modules" back to the calls to the CompileImage defines fixes the
> > regression. For kernels > 5.10 this is not needed, but it doesn't cause
> > any harm either.
> Why is >5.10 not affected? Can we backport the fix? I'd like to avoid
> adding extra unnecessary build step that slow down running make
> target/install.
>

You added 5ed1e5140a80558ab47fd70410ae3242bed5becf because of changes in
this area. I imagine one would need to look there.

I don't see adding "modules" back slowing anything down, because the
modules are already built when CompileImage is run. "modules" just makes
sure on kernels < 5.10 that MODPOST is run and the Module.symvers file
is complete.

> > Tested with kernels 5.4.x and 5.10.x.
> >
> > Signed-off-by: Sebastian Kemper <sebastian_ml at gmx.net>
> > ---
> >  include/kernel-defaults.mk | 9 +++++++--
> >  1 file changed, 7 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/kernel-defaults.mk b/include/kernel-defaults.mk
> > index 4b0b136a03..1b3b4497a2 100644
> > --- a/include/kernel-defaults.mk
> > +++ b/include/kernel-defaults.mk
> > @@ -147,12 +147,17 @@ define Kernel/CopyImage
> >  	}
> >  endef
> >
> > +# Always add "modules" so a proper Module.symvers file is written that
> > +# also contains symbols from the kernel modules. Without these symbols
> > +# external packages that depend on exported symbols from kernel modules
> > +# will fail to build.
> >  define Kernel/CompileImage/Default
> >  	rm -f $(TARGET_DIR)/init
> > -	+$(KERNEL_MAKE) $(KERNEL_MAKEOPTS_IMAGE) $(if $(KERNELNAME),$(KERNELNAME),all)
> > +	+$(KERNEL_MAKE) $(KERNEL_MAKEOPTS_IMAGE) $(if $(KERNELNAME),$(KERNELNAME),all) modules
> >  	$(call Kernel/CopyImage)
> >  endef
> >
> > +# Here as well, always add "modules", see comment above.
> >  ifneq ($(CONFIG_TARGET_ROOTFS_INITRAMFS),)
> >  define Kernel/CompileImage/Initramfs
> >  	$(call Kernel/Configure/Initramfs)
> > @@ -173,7 +178,7 @@ endif
> >  # ?	$(if $(CONFIG_TARGET_INITRAMFS_COMPRESSION_LZ4),)
> >  	$(if $(CONFIG_TARGET_INITRAMFS_COMPRESSION_ZSTD),$(STAGING_DIR_HOST)/bin/zstd -T0 -f -o $(KERNEL_BUILD_DIR)/initrd.cpio.zstd $(KERNEL_BUILD_DIR)/initrd.cpio)
> >  endif
> > -	+$(KERNEL_MAKE) $(KERNEL_MAKEOPTS_IMAGE) $(if $(KERNELNAME),$(KERNELNAME),all)
> > +	+$(KERNEL_MAKE) $(KERNEL_MAKEOPTS_IMAGE) $(if $(KERNELNAME),$(KERNELNAME),all) modules
> Why do this for initramfs as well?

Again, because Module.symvers is regenerated. And whatever
Module.symvers file is the last one to be generated will make its way
into the SDK.

>
> - Felix