Re: [tip: x86/core] x86/cpufeatures: Generate the <asm/cpufeaturemasks.h> header based on build config

[Jakub Kicinski]

On Sun, 23 Mar 2025 18:06:07 +0000 Michael Kelley wrote:
> > > With this commit, changing the modified time on .config causes cpufeaturemasks.h
> > > to be rebuilt, and then all dependencies on cpufeaturemasks.h are also rebuilt. It
> > > doesn't figure out that nothing has changed in .config. A fair amount of
> > > unnecessary rebuilding is done.  
> > 
> > Hmm. Normally it is the tools that *create* the config file that
> > should notice that it hasn't changed, and not write a new one.
> > 
> > See conf_write() in scripts/kconfig/confdata.c for the particular case
> > of the regular "make *config*'
> >   
> > > I'm seeing this changed behavior specifically with the make-kpkg script on
> > > Ubuntu 20.04. It's not clear why make-kpkg is updating the modified time
> > > on .config, but prior to this commit, that update had no negative effect.  
> > 
> > It sounds like the make-kpkg script should probably be fixed to not
> > overwrite the config file unconditionally, and do that "make
> > oldconfig" and friends do, which is to notice when it hasn't changed
> > and leave it alone.
> >   
> 
> For the record, the problem with make-kpkg is that it does "make syncconfig"
> multiple times in prepping to build, sometimes passing environment
> variable ARCH=x86 and sometimes ARCH=x86_64. scripts/kconfig/conf uses
> that environment variable to decorate the header of the generated .config
> file, and to force a rebuild in include/config/auto.conf.cmd if the value changes.
> So flipping between two different values causes a "new" .config to be
> generated even though nothing of substance has changed.
> 
> I hacked the make-kpkg scripts to consistently use ARCH=x86 and that
> solves the problem. But this is a private hack as make-kpkg hasn't been
> updated in years and is no longer supported. In Ubuntu releases, it is
> deprecated after 20.04.

FWIW this also makes netdev CI melt down. Combination of allmodconfig
insisting on WERROR, us insisting on W=1, and this change insisting on
full kernel rebuild every time .config is touched :( IOW we do:

	make allmodconfig
	./scripts/config -d werror
	make -j W=1

Could a hack like this fly? I don't see an easy way out ..

diff --git a/arch/x86/Makefile b/arch/x86/Makefile
index a607769a25af..35401572d039 100644
--- a/arch/x86/Makefile
+++ b/arch/x86/Makefile
@@ -276,7 +276,9 @@ cpufeaturemasks.awk := $(srctree)/arch/x86/tools/cpufeaturemasks.awk
 cpufeatures_hdr := $(srctree)/arch/x86/include/asm/cpufeatures.h
 targets += $(cpufeaturemasks.hdr)
 quiet_cmd_gen_featuremasks = GEN     $@
-      cmd_gen_featuremasks = $(AWK) -f $(cpufeaturemasks.awk) $(cpufeatures_hdr) $(KCONFIG_CONFIG) > $@
+      cmd_gen_featuremasks = $(AWK) -f $(cpufeaturemasks.awk) $(cpufeatures_hdr) $(KCONFIG_CONFIG) > $@.new; \
+       if [ -e $@ ] && cmp -s $@ $@.new; then rm $@.new; \
+       else mv $@.new $@; fi
 
 $(cpufeaturemasks.hdr): $(cpufeaturemasks.awk) $(cpufeatures_hdr) $(KCONFIG_CONFIG) FORCE
        $(shell mkdir -p $(dir $@))