On LoongArch, The "address" of a per-CPU variable symbol is actually an
offset from $r21.  The value is nearing the loading address of main
kernel image, but far from the address of modules.  We need to tell the
compiler that a PC-relative addressing with 32-bit offset is not
sufficient for local per-CPU variables.

After some discussion with GCC maintainers, we implemented this
attribute to indicate that a PC-relative addressing with 64-bit offset
should be used.

This attribute should be available in GCC 13 release.  Some early GCC 13
snapshots may support -mexplicit-relocs but lack this attribute, we
simply reject them.

Signed-off-by: Xi Ruoyao <xry111@xry111.site>
Link: https://gcc.gnu.org/r13-2199
---
 arch/loongarch/Makefile             | 3 +++
 arch/loongarch/include/asm/percpu.h | 8 ++++++++
 2 files changed, 11 insertions(+)

diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile
index 1563747c4fa8..593818a61741 100644
--- a/arch/loongarch/Makefile
+++ b/arch/loongarch/Makefile
@@ -53,6 +53,9 @@ LDFLAGS_vmlinux			+= -G0 -static -n -nostdlib
 # combination of a "new" assembler and "old" compiler is not supported.  Either
 # upgrade the compiler or downgrade the assembler.
 ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS
+ifeq ($(shell echo '__has_attribute(model)' | $(CC) -E -P - 2> /dev/null), 0)
+$(error "C compiler must support model attribute if using explicit relocs")
+endif
 cflags-y			+= -mexplicit-relocs
 else
 cflags-y			+= $(call cc-option,-mno-explicit-relocs)
diff --git a/arch/loongarch/include/asm/percpu.h b/arch/loongarch/include/asm/percpu.h
index 0bd6b0110198..dd7fcc553efa 100644
--- a/arch/loongarch/include/asm/percpu.h
+++ b/arch/loongarch/include/asm/percpu.h
@@ -8,6 +8,14 @@
 #include <asm/cmpxchg.h>
 #include <asm/loongarch.h>
 
+#if defined(MODULE) && defined(CONFIG_AS_HAS_EXPLICIT_RELOCS)
+/* The "address" (in fact, offset from $r21) of a per-CPU variable is close
+ * to the load address of main kernel image, but far from where the modules are
+ * loaded.  Tell the compiler this fact.
+ */
+# define PER_CPU_ATTRIBUTES __attribute__((model("extreme")))
+#endif
+
 /* Use r21 for fast access */
 register unsigned long __my_cpu_offset __asm__("$r21");
 
-- 
2.37.0

On Mon, 2022-08-29 at 21:31 +0800, Xi Ruoyao wrote:
> diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile
> index 1563747c4fa8..593818a61741 100644
> --- a/arch/loongarch/Makefile
> +++ b/arch/loongarch/Makefile
> @@ -53,6 +53,9 @@ LDFLAGS_vmlinux                       += -G0 -static -n -nostdlib
>  # combination of a "new" assembler and "old" compiler is not supported.  Either
>  # upgrade the compiler or downgrade the assembler.
>  ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS
> +ifeq ($(shell echo '__has_attribute(model)' | $(CC) -E -P - 2> /dev/null), 0)
> +$(error "C compiler must support model attribute if using explicit relocs")
> +endif

Self review:

I'm wondering if we really need this thing...  There won't be a GCC
version released with explicit relocation but without model attribute
(GCC 13 starts to support them both).

But without a check, if someone uses an early GCC 13 snapshot and
ignores the -Wattributes warning, the system will suddenly blow up
loading a module with per-CPU variable defined.

Maybe "-Werror=attributes" is better, but is it OK to add a -Werror=
option for entire Linux tree?

>  cflags-y                       += -mexplicit-relocs
>  else
>  cflags-y                       += $(call cc-option,-mno-explicit-relocs)

-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University

Hi, Ruoyao,

On Mon, Aug 29, 2022 at 11:02 PM Xi Ruoyao <xry111@xry111.site> wrote:
>
> On Mon, 2022-08-29 at 21:31 +0800, Xi Ruoyao wrote:
> > diff --git a/arch/loongarch/Makefile b/arch/loongarch/Makefile
> > index 1563747c4fa8..593818a61741 100644
> > --- a/arch/loongarch/Makefile
> > +++ b/arch/loongarch/Makefile
> > @@ -53,6 +53,9 @@ LDFLAGS_vmlinux                       += -G0 -static -n -nostdlib
> >  # combination of a "new" assembler and "old" compiler is not supported.  Either
> >  # upgrade the compiler or downgrade the assembler.
> >  ifdef CONFIG_AS_HAS_EXPLICIT_RELOCS
> > +ifeq ($(shell echo '__has_attribute(model)' | $(CC) -E -P - 2> /dev/null), 0)
> > +$(error "C compiler must support model attribute if using explicit relocs")
> > +endif
>
> Self review:
>
> I'm wondering if we really need this thing...  There won't be a GCC
> version released with explicit relocation but without model attribute
> (GCC 13 starts to support them both).
>
> But without a check, if someone uses an early GCC 13 snapshot and
> ignores the -Wattributes warning, the system will suddenly blow up
> loading a module with per-CPU variable defined.
>
> Maybe "-Werror=attributes" is better, but is it OK to add a -Werror=
> option for entire Linux tree?
I think we can remove it entirely, and then this patch seems can be
squashed into patch 2.

Huacai
>
> >  cflags-y                       += -mexplicit-relocs
> >  else
> >  cflags-y                       += $(call cc-option,-mno-explicit-relocs)
>
> --
> Xi Ruoyao <xry111@xry111.site>
> School of Aerospace Science and Technology, Xidian University
>