On Wed, Mar 22, 2017 at 1:24 PM, Dmitry Vyukov <dvyukov@google.com> wrote:
> Two problems was reported with READ_ONCE_NOCHECK in arch_atomic_read:
> 1. Andrey Ryabinin reported significant binary size increase
> (+400K of text). READ_ONCE_NOCHECK is intentionally compiled to
> non-inlined function call, and I counted 640 copies of it in my vmlinux.
> 2. Arnd Bergmann reported a new splat of too large frame sizes.
>
> A single inlined KASAN check is very cheap, a non-inlined function
> call with KASAN/KCOV instrumentation can easily be more expensive.
>
> Switch to READ_ONCE() in arch_atomic_read().
>
> Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
> Reported-by: Arnd Bergmann <arnd@arndb.de>
> Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Will Deacon <will.deacon@arm.com>
> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: linux-mm@kvack.org
> Cc: x86@kernel.org
> Cc: linux-kernel@vger.kernel.org
> Cc: kasan-dev@googlegroups.com
> ---
> arch/x86/include/asm/atomic.h | 15 ++++++---------
> 1 file changed, 6 insertions(+), 9 deletions(-)
>
> diff --git a/arch/x86/include/asm/atomic.h b/arch/x86/include/asm/atomic.h
> index 0cde164f058a..46e53bbf7ce3 100644
> --- a/arch/x86/include/asm/atomic.h
> +++ b/arch/x86/include/asm/atomic.h
> @@ -24,10 +24,13 @@
> static __always_inline int arch_atomic_read(const atomic_t *v)
> {
> /*
> - * We use READ_ONCE_NOCHECK() because atomic_read() contains KASAN
> - * instrumentation. Double instrumentation is unnecessary.
> + * Note: READ_ONCE() here leads to double instrumentation as
> + * both READ_ONCE() and atomic_read() contain instrumentation.
> + * This is deliberate choice. READ_ONCE_NOCHECK() is compiled to a
> + * non-inlined function call that considerably increases binary size
> + * and stack usage under KASAN.
> */
> - return READ_ONCE_NOCHECK((v)->counter);
> + return READ_ONCE((v)->counter);
> }
The change looks good, but the same one is needed in atomic64.h
Arnd
> Two problems was reported with READ_ONCE_NOCHECK in arch_atomic_read:
> 1. Andrey Ryabinin reported significant binary size increase
> (+400K of text). READ_ONCE_NOCHECK is intentionally compiled to
> non-inlined function call, and I counted 640 copies of it in my vmlinux.
> 2. Arnd Bergmann reported a new splat of too large frame sizes.
>
> A single inlined KASAN check is very cheap, a non-inlined function
> call with KASAN/KCOV instrumentation can easily be more expensive.
>
> Switch to READ_ONCE() in arch_atomic_read().
>
> Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
> Reported-by: Arnd Bergmann <arnd@arndb.de>
> Reported-by: Andrey Ryabinin <aryabinin@virtuozzo.com>
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Will Deacon <will.deacon@arm.com>
> Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: linux-mm@kvack.org
> Cc: x86@kernel.org
> Cc: linux-kernel@vger.kernel.org
> Cc: kasan-dev@googlegroups.com
> ---
> arch/x86/include/asm/atomic.h | 15 ++++++---------
> 1 file changed, 6 insertions(+), 9 deletions(-)
>
> diff --git a/arch/x86/include/asm/atomic.h b/arch/x86/include/asm/atomic.h
> index 0cde164f058a..46e53bbf7ce3 100644
> --- a/arch/x86/include/asm/atomic.h
> +++ b/arch/x86/include/asm/atomic.h
> @@ -24,10 +24,13 @@
> static __always_inline int arch_atomic_read(const atomic_t *v)
> {
> /*
> - * We use READ_ONCE_NOCHECK() because atomic_read() contains KASAN
> - * instrumentation. Double instrumentation is unnecessary.
> + * Note: READ_ONCE() here leads to double instrumentation as
> + * both READ_ONCE() and atomic_read() contain instrumentation.
> + * This is deliberate choice. READ_ONCE_NOCHECK() is compiled to a
> + * non-inlined function call that considerably increases binary size
> + * and stack usage under KASAN.
> */
> - return READ_ONCE_NOCHECK((v)->counter);
> + return READ_ONCE((v)->counter);
> }
The change looks good, but the same one is needed in atomic64.h
Arnd