Received-SPF: pass (google.com: domain of linux-kernel+bounces-108793-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Message-ID: <e8c4062df63f3e8bc8bb2d7209fa2a2a44bd7ed3.camel@xry111.site>
Subject: Re: [PATCH] LoongArch: Change __my_cpu_offset definition to avoid
 mis-optimization
From: Xi Ruoyao <xry111@xry111.site>
To: Huacai Chen <chenhuacai@loongson.cn>, Arnd Bergmann <arnd@arndb.de>, 
 Huacai Chen <chenhuacai@kernel.org>
Cc: loongarch@lists.linux.dev, linux-arch@vger.kernel.org, Xuefeng Li
	 <lixuefeng@loongson.cn>, Guo Ren <guoren@kernel.org>, Xuerui Wang
	 <kernel@xen0n.name>, Jiaxun Yang <jiaxun.yang@flygoat.com>, 
	linux-kernel@vger.kernel.org, loongson-kernel@lists.loongnix.cn, 
	stable@vger.kernel.org, Xiaotian Wu <wuxiaotian@loongson.cn>, Miao Wang
	 <shankerwangmiao@gmail.com>, Xing Li <lixing@loongson.cn>, Hongchen Zhang
	 <zhanghongchen@loongson.cn>, Rui Wang <wangrui@loongson.cn>
Date: Wed, 20 Mar 2024 18:27:26 +0800
In-Reply-To: <20240315024526.394772-1-chenhuacai@loongson.cn>
References: <20240315024526.394772-1-chenhuacai@loongson.cn>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.52.0 
Precedence: bulk
MIME-Version: 1.0

On Fri, 2024-03-15 at 10:45 +0800, Huacai Chen wrote:
> From GCC commit 3f13154553f8546a ("df-scan: remove ad-hoc handling of
> global regs in asms"), global registers will no longer be forced to add
> to the def-use chain. Then current_thread_info(), current_stack_pointer
> and __my_cpu_offset may be lifted out of the loop because they are no
> longer treated as "volatile variables".

Ooops...  I'm wondering why this issue has not blown up our systems
before.  The referred GCC commit is far before LoongArch CPUs are taped.

> This optimization is still correct for the current_thread_info() and
> current_stack_pointer usages because they are associated to a thread.
> However it is wrong for __my_cpu_offset because it is associated to a
> CPU rather than a thread: if the thread migrates to a different CPU in
> the loop, __my_cpu_offset should be changed.
>=20
> Change __my_cpu_offset definition to treat it as a "volatile variable",
> in order to avoid such a mis-optimization.
>=20
> Cc: stable@vger.kernel.org

I suppose we should add Fixes: 5b0b14e550a0 ("LoongArch: Add
atomic/locking header") here.

> Reported-by: Xiaotian Wu <wuxiaotian@loongson.cn>
> Reported-by: Miao Wang <shankerwangmiao@gmail.com>
> Signed-off-by: Xing Li <lixing@loongson.cn>
> Signed-off-by: Hongchen Zhang <zhanghongchen@loongson.cn>
> Signed-off-by: Rui Wang <wangrui@loongson.cn>
> Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
> ---
> =C2=A0arch/loongarch/include/asm/percpu.h | 6 +++++-
> =C2=A01 file changed, 5 insertions(+), 1 deletion(-)
>=20
> diff --git a/arch/loongarch/include/asm/percpu.h b/arch/loongarch/include=
/asm/percpu.h
> index 9b36ac003f89..03b98491d301 100644
> --- a/arch/loongarch/include/asm/percpu.h
> +++ b/arch/loongarch/include/asm/percpu.h
> @@ -29,7 +29,12 @@ static inline void set_my_cpu_offset(unsigned long off=
)
> =C2=A0	__my_cpu_offset =3D off;
> =C2=A0	csr_write64(off, PERCPU_BASE_KS);
> =C2=A0}
> -#define __my_cpu_offset __my_cpu_offset
> +
> +#define __my_cpu_offset					\
> +({							\
> +	__asm__ __volatile__("":"+r"(__my_cpu_offset));	\
> +	__my_cpu_offset;				\
> +})
> =C2=A0
> =C2=A0#define PERCPU_OP(op, asm_op, c_op)					\
> =C2=A0static __always_inline unsigned long __percpu_##op(void *ptr,		\

--=20
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University