Re: [PATCH] kbuild: prevent building single targets twice

[Mathias Krause <minipli@grsecurity.net>]

On 21.10.26 23:55, Nathan Chancellor wrote:
> [...]
> I forgot to mention, it looks like this is a known corner case in Kbuild
> that has come up before:
> 
>   https://lore.kernel.org/7b663b86-4df0-c3df-18a7-e4f47f900be4@kernel.org/
>   https://lore.kernel.org/CAK7LNAQhORSPEuB=XFa_i5EcQfsJ6uXy0MM5YWo_84eLGcoxNQ@mail.gmail.com/

Yeah, I saw these too when looking for related issues and fixes.
However, even Masahiro admitted, that this is just working around the
issue instead of solving the more fundamental kbuild bug.

> 
> This patch might be the fix that Masahiro could not figure out but it
> might be better to just forsake drivers-y here and use obj-y properly,
> especially since it looks like core-y is not even used for x86 anymore
> (but maybe these objects still need to linked towards the end of the
> kernel?). This diff appears to resolve the reported issue but I did not
> try running the result.
> 
> diff --git a/arch/x86/Kbuild b/arch/x86/Kbuild
> index 36b985d0e7bf..bd0621207c0c 100644
> --- a/arch/x86/Kbuild
> +++ b/arch/x86/Kbuild
> @@ -36,5 +36,12 @@ obj-$(CONFIG_KEXEC_FILE) += purgatory/
>  
>  obj-y += virt/
>  
> +obj-$(CONFIG_PCI) += pci/
> +
> +# suspend and hibernation support
> +obj-$(CONFIG_PM) += power/
> +
> +obj-$(CONFIG_VIDEO) += video/
> +
>  # for cleaning
>  subdir- += boot tools
> diff --git a/arch/x86/Makefile b/arch/x86/Makefile
> index b7e08519b8c1..9a872bd9520e 100644
> --- a/arch/x86/Makefile
> +++ b/arch/x86/Makefile
> @@ -279,14 +279,6 @@ archprepare: $(cpufeaturemasks.hdr)
>  
>  libs-y  += arch/x86/lib/
>  
> -# drivers-y are linked after core-y
> -drivers-$(CONFIG_PCI)            += arch/x86/pci/
> -
> -# suspend and hibernation support
> -drivers-$(CONFIG_PM) += arch/x86/power/
> -
> -drivers-$(CONFIG_VIDEO) += arch/x86/video/
> -
>  ####
>  # boot loader support. Several targets are kept for legacy purposes
>  

This only "fixes" (read on why in quotes) this one instance but leaves
others unsolved.

Already visible in the diff context is 'libs-y' which has the same issue
and a real behavioral change, depending on which compiler invocation
runs last for arch/x86/lib/delay.o, it'll either be compiled with or
without -pg under KCSAN.

Others are:
- arch/x86/platform/pvh/enlighten.o (via arch/x86/Kbuild),
- drivers/char/ipmi/ipmi_msghandler.o (via drivers/Makefile,
  intentionally there for init ordering),
- drivers/input/serio/i8042.o (also via drivers/Makefile)

But what's worse, your change also affects link order which also defines
runtime initialization order. I don't want to crawl into that rabbit
hole and look for all the implicit dependencies that'd be broken by that
change...

One obvious, however, would be pci_subsys_init() which would then run
before acpi_init(), leaving out ACPI-discovered PCI roots from getting
discovered.

So no, I don't like your change and prefer the more fundamental kbuild
fix to get this sorted out once for all.

Thanks,
Mathias