linux-next: build problems (Was: [PATCH v3 1/2] Provide READ_ONCE_NOCHECK())

linux-next: build problems (Was: [PATCH v3 1/2] Provide READ_ONCE_NOCHECK())

[Stephen Rothwell]

Hi Andrey,

On Tue, 13 Oct 2015 18:28:07 +0300 Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:
>
> Some code may perform racy by design memory reads. This could be harmless,
> yet such code may produce KASAN warnings.
> 
> To hide such accesses from KASAN this patch introduces READ_ONCE_NOCHECK()
> macro. KASAN will not check the memory accessed by READ_ONCE_NOCHECK().
> 
> This patch creates __read_once_size_nocheck() a clone of
> __read_once_size_check() (renamed __read_once_size()).
> The only difference between them is 'no_sanitized_address' attribute
> appended to '*_nocheck' function. This attribute tells the compiler that
> instrumentation of memory accesses should not be applied to that function.
> We declare it as static '__maybe_unsed' because GCC is not capable to
> inline such function: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67368
> 
> With KASAN=n READ_ONCE_NOCHECK() is just a clone of READ_ONCE().
> 
> Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
> ---
>  include/linux/compiler-gcc.h | 13 ++++++++++
>  include/linux/compiler.h     | 60 ++++++++++++++++++++++++++++++++++----------
>  2 files changed, 60 insertions(+), 13 deletions(-)

I am pretty sure that this patch is causing quite a bit of compile
breakage in linux-next today.  During the day I compile with gcc 4.9.0
and did not see any problems with c86_64 allmodconfig, or i386
defconfig etc, but overnight we compile with older compilers (gcc 4.6.3
in particular) and are getting quite a few errors:

>From an i386 allnoconfig build:

arch/x86/entry/vdso/vdso32.so.dbg: undefined symbols found
/home/kisskb/slave/src/arch/x86/entry/vdso/Makefile:154: recipe for target 'arch/x86/entry/vdso/vdso32.so.dbg' failed

>From an x86_64 allnoconfig build:

arch/x86/entry/vdso/vclock_gettime.o: In function `__read_once_size_check':
vclock_gettime.c:(.text+0x5f): undefined reference to `memcpy'
arch/x86/entry/vdso/vgetcpu.o: In function `__read_once_size_check':
vgetcpu.c:(.text+0x2f): undefined reference to `memcpy'

and several others ...
-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au

Re: linux-next: build problems (Was: [PATCH v3 1/2] Provide READ_ONCE_NOCHECK())

[Andrey Ryabinin]

On 10/15/2015 12:18 PM, Stephen Rothwell wrote:
> Hi Andrey,
> 
> On Tue, 13 Oct 2015 18:28:07 +0300 Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:
>>
>> Some code may perform racy by design memory reads. This could be harmless,
>> yet such code may produce KASAN warnings.
>>
>> To hide such accesses from KASAN this patch introduces READ_ONCE_NOCHECK()
>> macro. KASAN will not check the memory accessed by READ_ONCE_NOCHECK().
>>
>> This patch creates __read_once_size_nocheck() a clone of
>> __read_once_size_check() (renamed __read_once_size()).
>> The only difference between them is 'no_sanitized_address' attribute
>> appended to '*_nocheck' function. This attribute tells the compiler that
>> instrumentation of memory accesses should not be applied to that function.
>> We declare it as static '__maybe_unsed' because GCC is not capable to
>> inline such function: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67368
>>
>> With KASAN=n READ_ONCE_NOCHECK() is just a clone of READ_ONCE().
>>
>> Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
>> ---
>>  include/linux/compiler-gcc.h | 13 ++++++++++
>>  include/linux/compiler.h     | 60 ++++++++++++++++++++++++++++++++++----------
>>  2 files changed, 60 insertions(+), 13 deletions(-)
> 
> I am pretty sure that this patch is causing quite a bit of compile
> breakage in linux-next today.  During the day I compile with gcc 4.9.0
> and did not see any problems with c86_64 allmodconfig, or i386
> defconfig etc, but overnight we compile with older compilers (gcc 4.6.3
> in particular) and are getting quite a few errors:
> 

Looks like that older GCC doesn't like __alias (or combination of static __always_inline __alias).
It creates outline and unused copy of __read_once_size_check() function in the object file.
Should be easy to work around this.

> From an i386 allnoconfig build:
> 
> arch/x86/entry/vdso/vdso32.so.dbg: undefined symbols found
> /home/kisskb/slave/src/arch/x86/entry/vdso/Makefile:154: recipe for target 'arch/x86/entry/vdso/vdso32.so.dbg' failed
> 
> From an x86_64 allnoconfig build:
> 
> arch/x86/entry/vdso/vclock_gettime.o: In function `__read_once_size_check':
> vclock_gettime.c:(.text+0x5f): undefined reference to `memcpy'
> arch/x86/entry/vdso/vgetcpu.o: In function `__read_once_size_check':
> vgetcpu.c:(.text+0x2f): undefined reference to `memcpy'
> 
> and several others ...
> 

[PATCH] compiler, READ_ONCE: Fix build failure with some older GCC

[Andrey Ryabinin]

Some old versions of GCC doesn't handle __alias (or maybe a combination
of 'static __always_inline __alias') right.

GCC creates outline and unused copy of __read_once_size_check()
function in the object file which references memcpy and causes
the build failure:
	arch/x86/entry/vdso/vclock_gettime.o: In function `__read_once_size_check':
	vclock_gettime.c:(.text+0x5f): undefined reference to `memcpy'
	arch/x86/entry/vdso/vgetcpu.o: In function `__read_once_size_check':
	vgetcpu.c:(.text+0x2f): undefined reference to `memcpy'

We could avoid using alias to work around this problem.

Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
---
 include/linux/compiler.h | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/linux/compiler.h b/include/linux/compiler.h
index aa2ae4c..3436a4c 100644
--- a/include/linux/compiler.h
+++ b/include/linux/compiler.h
@@ -231,8 +231,11 @@ void __read_once_size_nocheck(const volatile void *p, void *res, int size)
 	__READ_ONCE_SIZE;
 }
 #else
-static __always_inline __alias(__read_once_size_check)
-void __read_once_size_nocheck(const volatile void *p, void *res, int size);
+static __always_inline
+void __read_once_size_nocheck(const volatile void *p, void *res, int size)
+{
+	__READ_ONCE_SIZE;
+}
 #endif
 
 static __always_inline void __write_once_size(volatile void *p, void *res, int size)
-- 
2.4.9

Re: [PATCH] compiler, READ_ONCE: Fix build failure with some older GCC

[Ingo Molnar]

* Andrey Ryabinin <aryabinin@virtuozzo.com> wrote:

> Some old versions of GCC doesn't handle __alias (or maybe a combination
> of 'static __always_inline __alias') right.
> 
> GCC creates outline and unused copy of __read_once_size_check()
> function in the object file which references memcpy and causes
> the build failure:
> 	arch/x86/entry/vdso/vclock_gettime.o: In function `__read_once_size_check':
> 	vclock_gettime.c:(.text+0x5f): undefined reference to `memcpy'
> 	arch/x86/entry/vdso/vgetcpu.o: In function `__read_once_size_check':
> 	vgetcpu.c:(.text+0x2f): undefined reference to `memcpy'
> 
> We could avoid using alias to work around this problem.
> 
> Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
> Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
> ---
>  include/linux/compiler.h | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/compiler.h b/include/linux/compiler.h
> index aa2ae4c..3436a4c 100644
> --- a/include/linux/compiler.h
> +++ b/include/linux/compiler.h
> @@ -231,8 +231,11 @@ void __read_once_size_nocheck(const volatile void *p, void *res, int size)
>  	__READ_ONCE_SIZE;
>  }
>  #else
> -static __always_inline __alias(__read_once_size_check)
> -void __read_once_size_nocheck(const volatile void *p, void *res, int size);
> +static __always_inline
> +void __read_once_size_nocheck(const volatile void *p, void *res, int size)
> +{
> +	__READ_ONCE_SIZE;
> +}
>  #endif

Yeah, so could you please re-send the original two patches, with all the changes 
(typo fixes, renaming, this build fix, etc.) propagated that were found in testing 
and that were suggested in the review thread?

I'll remove the broken commits from linux-next meanwhile.

Thanks,

	Ingo