[PATCH net-next] sh: add missing fpscr_values

[PATCH net-next] sh: add missing fpscr_values

[Alexei Starovoitov]

GCC on SuperH for the following program:
int foo(int i)
{
	static const void *table[2] = {
		[0] = &&label1,
		[1] = &&label2,
	};
	goto *table[i & 1];

	label1:
		return 1;
	label2:
		return 2;
}
generates code:
	mov.l   .L9,r1
	mov     #1,r0
	rts
	lds.l   @r1+,fpscr
.L9:
	.long   __fpscr_values+4

so kernel has to provide body of fpscr_values just like glibc does

Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Cc: Daniel Borkmann <dborkman@redhat.com>
---

kernel modules can hit the same problem, so export it

 arch/sh/lib/Makefile       |    2 +-
 arch/sh/lib/fpscr_values.c |    4 ++++
 2 files changed, 5 insertions(+), 1 deletion(-)
 create mode 100644 arch/sh/lib/fpscr_values.c

diff --git a/arch/sh/lib/Makefile b/arch/sh/lib/Makefile
index 3baff31e58cf..0d0c7ae21cfe 100644
--- a/arch/sh/lib/Makefile
+++ b/arch/sh/lib/Makefile
@@ -8,7 +8,7 @@ lib-y  = delay.o memmove.o memchr.o \
 # Extracted from libgcc
 obj-y += movmem.o ashldi3.o ashrdi3.o lshrdi3.o \
 	 ashlsi3.o ashrsi3.o ashiftrt.o lshrsi3.o \
-	 udiv_qrnnd.o
+	 udiv_qrnnd.o fpscr_values.o
 
 udivsi3-y			:= udivsi3_i4i-Os.o
 
diff --git a/arch/sh/lib/fpscr_values.c b/arch/sh/lib/fpscr_values.c
new file mode 100644
index 000000000000..58153272aba1
--- /dev/null
+++ b/arch/sh/lib/fpscr_values.c
@@ -0,0 +1,4 @@
+#include <linux/module.h>
+
+long __fpscr_values[2] = { 0L, 0x80000L };
+EXPORT_SYMBOL(__fpscr_values);
-- 
1.7.9.5

Re: [PATCH net-next] sh: add missing fpscr_values

[Daniel Borkmann]

On 04/02/2014 04:54 AM, Alexei Starovoitov wrote:
> GCC on SuperH for the following program:
> int foo(int i)
> {
> 	static const void *table[2] = {
> 		[0] = &&label1,
> 		[1] = &&label2,
> 	};
> 	goto *table[i & 1];
>
> 	label1:
> 		return 1;
> 	label2:
> 		return 2;
> }
> generates code:
> 	mov.l   .L9,r1
> 	mov     #1,r0
> 	rts
> 	lds.l   @r1+,fpscr
> .L9:
> 	.long   __fpscr_values+4
>
> so kernel has to provide body of fpscr_values just like glibc does
>
> Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
> Reported-by: Fengguang Wu <fengguang.wu@intel.com>
> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
> Cc: Daniel Borkmann <dborkman@redhat.com>

Acked-by: Daniel Borkmann <dborkman@redhat.com>

Re: [PATCH net-next] sh: add missing fpscr_values

[Geert Uytterhoeven]

On Wed, Apr 2, 2014 at 4:54 AM, Alexei Starovoitov <ast@plumgrid.com> wrote:
> GCC on SuperH for the following program:
> int foo(int i)
> {
>         static const void *table[2] = {
>                 [0] = &&label1,
>                 [1] = &&label2,
>         };
>         goto *table[i & 1];
>
>         label1:
>                 return 1;
>         label2:
>                 return 2;
> }
> generates code:
>         mov.l   .L9,r1
>         mov     #1,r0
>         rts
>         lds.l   @r1+,fpscr
> .L9:
>         .long   __fpscr_values+4
>
> so kernel has to provide body of fpscr_values just like glibc does
>
> Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
> Reported-by: Fengguang Wu <fengguang.wu@intel.com>
> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
> Cc: Daniel Borkmann <dborkman@redhat.com>
> ---
>
> kernel modules can hit the same problem, so export it
>
>  arch/sh/lib/Makefile       |    2 +-
>  arch/sh/lib/fpscr_values.c |    4 ++++
>  2 files changed, 5 insertions(+), 1 deletion(-)
>  create mode 100644 arch/sh/lib/fpscr_values.c
>
> diff --git a/arch/sh/lib/Makefile b/arch/sh/lib/Makefile
> index 3baff31e58cf..0d0c7ae21cfe 100644
> --- a/arch/sh/lib/Makefile
> +++ b/arch/sh/lib/Makefile
> @@ -8,7 +8,7 @@ lib-y  = delay.o memmove.o memchr.o \
>  # Extracted from libgcc
>  obj-y += movmem.o ashldi3.o ashrdi3.o lshrdi3.o \
>          ashlsi3.o ashrsi3.o ashiftrt.o lshrsi3.o \
> -        udiv_qrnnd.o
> +        udiv_qrnnd.o fpscr_values.o
>
>  udivsi3-y                      := udivsi3_i4i-Os.o
>
> diff --git a/arch/sh/lib/fpscr_values.c b/arch/sh/lib/fpscr_values.c
> new file mode 100644
> index 000000000000..58153272aba1
> --- /dev/null
> +++ b/arch/sh/lib/fpscr_values.c
> @@ -0,0 +1,4 @@
> +#include <linux/module.h>
> +
> +long __fpscr_values[2] = { 0L, 0x80000L };
> +EXPORT_SYMBOL(__fpscr_values);

Pardon my ignorance, but according to
http://www.sourceware.org/ml/libc-alpha/2003-06/msg00027.html
__fpscr_values is needed for FPU support only.

The kernel is built with -m4-nofpu. If this option is used, gcc doesn't generate
the reference to __fpscr_values?

Why do we need symbols for userspace FPU support in the kernel?

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Re: [PATCH net-next] sh: add missing fpscr_values

[Geert Uytterhoeven]

On Wed, Apr 2, 2014 at 9:38 AM, Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> On Wed, Apr 2, 2014 at 4:54 AM, Alexei Starovoitov <ast@plumgrid.com> wrote:
>> GCC on SuperH for the following program:
>> int foo(int i)
>> {
>>         static const void *table[2] = {
>>                 [0] = &&label1,
>>                 [1] = &&label2,
>>         };
>>         goto *table[i & 1];
>>
>>         label1:
>>                 return 1;
>>         label2:
>>                 return 2;
>> }
>> generates code:
>>         mov.l   .L9,r1
>>         mov     #1,r0
>>         rts
>>         lds.l   @r1+,fpscr
>> .L9:
>>         .long   __fpscr_values+4
>>
>> so kernel has to provide body of fpscr_values just like glibc does
>>
>> Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
>> Reported-by: Fengguang Wu <fengguang.wu@intel.com>
>> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
>> Cc: Daniel Borkmann <dborkman@redhat.com>
>> ---
>>
>> kernel modules can hit the same problem, so export it
>>
>>  arch/sh/lib/Makefile       |    2 +-
>>  arch/sh/lib/fpscr_values.c |    4 ++++
>>  2 files changed, 5 insertions(+), 1 deletion(-)
>>  create mode 100644 arch/sh/lib/fpscr_values.c
>>
>> diff --git a/arch/sh/lib/Makefile b/arch/sh/lib/Makefile
>> index 3baff31e58cf..0d0c7ae21cfe 100644
>> --- a/arch/sh/lib/Makefile
>> +++ b/arch/sh/lib/Makefile
>> @@ -8,7 +8,7 @@ lib-y  = delay.o memmove.o memchr.o \
>>  # Extracted from libgcc
>>  obj-y += movmem.o ashldi3.o ashrdi3.o lshrdi3.o \
>>          ashlsi3.o ashrsi3.o ashiftrt.o lshrsi3.o \
>> -        udiv_qrnnd.o
>> +        udiv_qrnnd.o fpscr_values.o
>>
>>  udivsi3-y                      := udivsi3_i4i-Os.o
>>
>> diff --git a/arch/sh/lib/fpscr_values.c b/arch/sh/lib/fpscr_values.c
>> new file mode 100644
>> index 000000000000..58153272aba1
>> --- /dev/null
>> +++ b/arch/sh/lib/fpscr_values.c
>> @@ -0,0 +1,4 @@
>> +#include <linux/module.h>
>> +
>> +long __fpscr_values[2] = { 0L, 0x80000L };
>> +EXPORT_SYMBOL(__fpscr_values);
>
> Pardon my ignorance, but according to
> http://www.sourceware.org/ml/libc-alpha/2003-06/msg00027.html
> __fpscr_values is needed for FPU support only.
>
> The kernel is built with -m4-nofpu. If this option is used, gcc doesn't generate
> the reference to __fpscr_values?
>
> Why do we need symbols for userspace FPU support in the kernel?

Ah, so this comes from these build failures in -next:
http://kisskb.ellerman.id.au/kisskb/buildresult/10875446/
http://kisskb.ellerman.id.au/kisskb/buildresult/10875446/

Indeed, SH2A-targets fail when built with a sh4-linux toolchain, as
sh4-linux-gcc doesn't understand -m2a-nofpu, so it's left out:

arch/sh/Makefile:
cflags-$(CONFIG_CPU_SH2A)               += $(call cc-option,-m2a,) \
                                           $(call cc-option,-m2a-nofpu,)

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Re: [PATCH net-next] sh: add missing fpscr_values

[David Miller]

From: Alexei Starovoitov <ast@plumgrid.com>
Date: Tue,  1 Apr 2014 19:54:04 -0700

> GCC on SuperH for the following program:
> int foo(int i)
> {
> 	static const void *table[2] = {
> 		[0] = &&label1,
> 		[1] = &&label2,
> 	};
> 	goto *table[i & 1];
> 
> 	label1:
> 		return 1;
> 	label2:
> 		return 2;
> }
> generates code:
> 	mov.l   .L9,r1
> 	mov     #1,r0
> 	rts
> 	lds.l   @r1+,fpscr
> .L9:
> 	.long   __fpscr_values+4
> 
> so kernel has to provide body of fpscr_values just like glibc does
> 
> Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
> Reported-by: Fengguang Wu <fengguang.wu@intel.com>
> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>

You guys really need to just fix the compiler, sorry.

Re: [PATCH net-next] sh: add missing fpscr_values

[Alexei Starovoitov]

On Wed, Apr 2, 2014 at 5:47 AM, David Miller <davem@davemloft.net> wrote:
> From: Alexei Starovoitov <ast@plumgrid.com>
> Date: Tue,  1 Apr 2014 19:54:04 -0700
>
>> GCC on SuperH for the following program:
>> int foo(int i)
>> {
>>       static const void *table[2] = {
>>               [0] = &&label1,
>>               [1] = &&label2,
>>       };
>>       goto *table[i & 1];
>>
>>       label1:
>>               return 1;
>>       label2:
>>               return 2;
>> }
>> generates code:
>>       mov.l   .L9,r1
>>       mov     #1,r0
>>       rts
>>       lds.l   @r1+,fpscr
>> .L9:
>>       .long   __fpscr_values+4
>>
>> so kernel has to provide body of fpscr_values just like glibc does
>>
>> Fixes: bd4cf0ed331a ("net: filter: rework/optimize internal BPF interpreter's instruction set")
>> Reported-by: Fengguang Wu <fengguang.wu@intel.com>
>> Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
>
> You guys really need to just fix the compiler, sorry.

agree.

To summarize:
according to Kaz:
"For sh2a, configuring gcc with --with-multilib-list=m2a,m2a-nofpu
 and specifying -m2a-nofpu during compiling kernel would work."

and as Geert pointed out we already have:
arch/sh/Makefile:
cflags-$(CONFIG_CPU_SH2A)               += $(call cc-option,-m2a,) \
                                           $(call cc-option,-m2a-nofpu,)

so no kernel changes needed and gcc just needs to be reconfigured.