Re: [PATCH v2 0/3] Detect changed compiler dependencies for full rebuild

[Kees Cook <kees@kernel.org>]

On Fri, May 09, 2025 at 01:44:09AM +0900, Masahiro Yamada wrote:
> On Sun, May 4, 2025 at 2:37 AM Kees Cook <kees@kernel.org> wrote:
> >
> > On Sat, May 03, 2025 at 06:39:28PM +0900, Masahiro Yamada wrote:
> > > On Sat, May 3, 2025 at 7:54 AM Kees Cook <kees@kernel.org> wrote:
> > > >
> > > >  v2:
> > > >   - switch from -include to -I with a -D gated include compiler-version.h
> > > >  v1: https://lore.kernel.org/lkml/20250501193839.work.525-kees@kernel.org/
> > >
> > >
> > > What do you think of my patch as a prerequisite?
> > > https://lore.kernel.org/linux-kbuild/20250503084145.1994176-1-masahiroy@kernel.org/T/#u
> > > Perhaps, can you implement this series more simply?
> > >
> > > My idea is to touch a single include/generated/global-rebuild.h
> > > rather than multiple files such as gcc-plugins-deps.h, integer-wrap.h, etc.
> > >
> > > When the file is touched, the entire kernel source tree will be rebuilt.
> > > This may rebuild more than needed (e.g. vdso) but I do not think
> > > it is a big deal.
> >
> > This is roughly where I started when trying to implement this, but I
> > didn't like the ergonomics of needing to scatter "touch" calls all over,
> > which was especially difficult for targets that shared a build rule but
> > may not all need to trigger a global rebuild. But what ultimately pushed
> > me away from it was when I needed to notice if a non-built source file
> > changed (the Clang .scl file), and I saw that I need to be dependency
> > driven rather than target driven. (Though perhaps there is a way to
> > address this with your global-rebuild.h?)
> >
> > As far as doing a full rebuild, if it had been available last week, I
> > probably would have used it, but now given the work that Nicolas, you,
> > and I have put into this, we have a viable way (I think) to make this
> > more specific. It does end up being a waste of time/resources to rebuild
> > stuff that doesn't need to be (efi-stub, vdso, boot code, etc), and that
> > does add up when I'm iterating on something that keeps triggering a full
> > rebuild. We already have to do the argument filtering for targets that
> > don't want randstruct, etc, so why not capitalize on that and make the
> > rebuild avoid those files too?
> 
> 
> efi-stub, vdso are very small.
> 
> Unless this turns out to be painful, I prefer
> a simpler implementation.
> 
> You will see how .scl file is handled.
> 
> See the below code:
> 
> 
> diff --git a/Kbuild b/Kbuild
> index f327ca86990c..85747239314c 100644
> --- a/Kbuild
> +++ b/Kbuild
> @@ -67,10 +67,20 @@ targets += $(atomic-checks)
>  $(atomic-checks): $(obj)/.checked-%: include/linux/atomic/%  FORCE
>         $(call if_changed,check_sha1)
> 
> +rebuild-$(CONFIG_GCC_PLUGINS)          += $(addprefix
> scripts/gcc-plugins/, $(GCC_PLUGIN))
> +rebuild-$(CONFIG_RANDSTRUCT)           += include/generated/randstruct_hash.h

These are in $(objtree)

> +rebuild-$(CONFIG_UBSAN_INTEGER_WRAP)   += scripts/integer-wrap-ignore.scl

This is in $(srctree)

> +
> +quiet_cmd_touch = TOUCH   $@
> +      cmd_touch = touch $@
> +
> +include/generated/global-rebuild.h: $(rebuild-y)
> +       $(call cmd,touch)

Is this rule going to find the right versions of the dependencies?

> --- a/Makefile
> +++ b/Makefile
> @@ -558,7 +558,8 @@ USERINCLUDE    := \
>                 -I$(srctree)/include/uapi \
>                 -I$(objtree)/include/generated/uapi \
>                  -include $(srctree)/include/linux/compiler-version.h \
> -                -include $(srctree)/include/linux/kconfig.h
> +                -include $(srctree)/include/linux/kconfig.h \
> +                -include $(objtree)/include/generated/global-rebuild.h

Instead of adding a new file, why not just touch compiler-version.h?

But whatever the case, sure, I can live with this. :)

-Kees

-- 
Kees Cook