Re: [PATCH 1/2] kbuild: remove cc-option test of -fno-stack-protector

[Nick Desaulniers]

On Mon, Jun 29, 2020 at 11:26 AM Nick Desaulniers
<ndesaulniers@xxxxxxxxxx> wrote:
>
> On Sat, Jun 27, 2020 at 4:59 AM Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> >
> > On Sat, Jun 27, 2020 at 5:13 AM Nick Desaulniers
> > <ndesaulniers@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Jun 26, 2020 at 12:00 PM Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
> > > >
> > > > diff --git a/arch/x86/kernel/cpu/Makefile b/arch/x86/kernel/cpu/Makefile
> > > > index dba6a83bc349..93792b457b81 100644
> > > > --- a/arch/x86/kernel/cpu/Makefile
> > > > +++ b/arch/x86/kernel/cpu/Makefile
> > > > @@ -17,8 +17,7 @@ KCOV_INSTRUMENT_perf_event.o := n
> > > >  KCSAN_SANITIZE_common.o := n
> > > >
> > > >  # Make sure load_percpu_segment has no stackprotector
> > > > -nostackp := $(call cc-option, -fno-stack-protector)
> > > > -CFLAGS_common.o                := $(nostackp)
> > > > +CFLAGS_common.o                := -fno-stack-protector
> > >
> > > Any time I see `:=` assignment to a CFLAGS variable, it's a red flag
> > > for overwriting existing CFLAGS, which is a common source of bugs.  I
> > > recognize the kernel is current a mix and match for:
> > >
> > > CFLAGS_<file>.o
> > >
> > > rules to either use `+=` or `:=`, but I wish we were consistent, and
> > > consistent in our use of `+=`.  For those rules, is there a difference
> > > between the use of `+=` and `:=` like there is for the more general
> > > case of appending to KBUILD_CFLAGS?  If not, it's ok to match the
> > > existing style, but it's curious to me in this patch to see a mixed
> > > use of `+=` and `:=`.
> >
> >
> > I think Kees mostly answered your question.
> >
> > Let me add some comments.
> >
> >
> > '+=' is the most used in kernel Makefiles, but
> > ':=' and '=' are also used.
> >
> > So, you are right, we are inconsistent.
> > This applies to not only CFLAGS_<file>.o, but also obj-y, etc.
> >
> > For example,
> > https://github.com/torvalds/linux/blob/v5.7/arch/arm64/kernel/Makefile#L15
> > 'obj-y :=' works since it is the first assignment to obj-y in that file.
> > 'obj-y +=' also works, of course.
> >
> > We can consistently use '+=' everywhere, but I do not send
> > patches for churn.
> >
> >
> > You can use any assignment operator to CFLAGS_<file>.o
> > if it is the first assignment in the Makefile.
> > Using '+=' is robust for future code insertion/removal, though.
> >
> >
> > If the right-hand side contains variable references,
> > there is important difference in the behavior.
> >
> > You may know two flavors in variables
> > (https://www.gnu.org/software/make/manual/make.html#Flavors)
>
> Cool, thanks for all the info.  With that, I'm happy with this patch.
>
> Reviewed-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
>
>
> > CFLAGS_foo.o := $(call cc-option,-fno-stack-protector)
> > The cc-option is expanded immediately when this line is parsed.
> > (So, the compiler is invoked for 'make clean' too)
> >
> >
> > CFLAGS_foo.o += $(call cc-option,-fno-stack-protector)
> > If this is the first assignment in the file,
> > '+=' act as '=', so the evaluation of cc-option
> > is delayed until $(CFLAGS_foo.o) is expanded.
> > (So, the compiler is NOT invoked for 'make clean')
>
> Ah, I think that may explain: I've been seeing the occasional warning
> from $(NM) when running `make clean` for ARCH=arm, I'll bet that's
> where this is coming from then.  Next time I reproduce it, I'll try to
> find maybe where we're using `:=` or `=` with `$(NM)`.
>
> Maybe arch/arm/boot/compressed/Makefile, KBSS_SZ is evaluated for
> `make clean`? (If you start an arm build, but kill it before vmlinux
> is created, then `make clean` I suspect that KBSS_SZ is evaluated?)

$ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make LLVM=1 -j71
$ rm vmlinux
$ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make LLVM=1 -j71 clean
llvm-nm: error: arch/arm/boot/compressed/../../../../vmlinux: No such
file or directory.

I suspect I see this from doing an ARCH=arm build, then building a
different ARCH and killing the build before vmlinux is produced or
`make clean` for a different ARCH, then run a `make clean` for
ARCH=arm.

Is the above (regarding lazy evaluation) not true for LDFLAGS_vmlinux?
I would have thought this would work:

diff --git a/arch/arm/boot/compressed/Makefile
b/arch/arm/boot/compressed/Makefile
index 00602a6fba04..9e7343ccc279 100644
--- a/arch/arm/boot/compressed/Makefile
+++ b/arch/arm/boot/compressed/Makefile
@@ -113,10 +113,9 @@ ccflags-y := -fpic $(call
cc-option,-mno-single-pic-base,) -fno-builtin \
 asflags-y := -DZIMAGE

 # Supply kernel BSS size to the decompressor via a linker symbol.
-KBSS_SZ = $(shell echo $$(($$($(NM) $(obj)/../../../../vmlinux | \
+LDFLAGS_vmlinux = --defsym _kernel_bss_size=$(shell echo $$(($$($(NM)
$(obj)/../../../../vmlinux | \
                sed -n -e 's/^\([^ ]*\) [AB] __bss_start$$/-0x\1/p' \
-                      -e 's/^\([^ ]*\) [AB] __bss_stop$$/+0x\1/p') )) )
-LDFLAGS_vmlinux = --defsym _kernel_bss_size=$(KBSS_SZ)
+                                        -e 's/^\([^ ]*\) [AB]
__bss_stop$$/+0x\1/p') )) )

-- 
Thanks,
~Nick Desaulniers