Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [RFC PATCH v2 07/10] lib: vdso: don't use READ_ONCE() in
 __c_kernel_time()
To:     Thomas Gleixner <tglx@linutronix.de>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Paul Mackerras <paulus@samba.org>,
        Michael Ellerman <mpe@ellerman.id.au>, arnd@arndb.de,
        vincenzo.frascino@arm.com, luto@kernel.org
Cc:     linux-kernel@vger.kernel.org, linuxppc-dev@lists.ozlabs.org,
        linux-arm-kernel@lists.infradead.org, linux-mips@vger.kernel.org,
        x86@kernel.org
References: <cover.1577111363.git.christophe.leroy@c-s.fr>
 <fc1ff722c7cbe63a63ae02ade3a714d2049d54a5.1577111367.git.christophe.leroy@c-s.fr>
 <87lfqfrp7d.fsf@nanos.tec.linutronix.de>
 <a995445f-9b00-ca13-d23a-1aea3b345718@c-s.fr>
 <878smes13d.fsf@nanos.tec.linutronix.de>
From:   Christophe Leroy <christophe.leroy@c-s.fr>
Message-ID: <dfd30b00-5d5c-29c3-6b91-aac0533635ca@c-s.fr>
Date:   Mon, 13 Jan 2020 07:52:31 +0100
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:60.0) Gecko/20100101
 Thunderbird/60.9.1
MIME-Version: 1.0
In-Reply-To: <878smes13d.fsf@nanos.tec.linutronix.de>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: fr
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


Le 11/01/2020 à 12:07, Thomas Gleixner a écrit :
> Christophe Leroy <christophe.leroy@c-s.fr> writes:
>>
>> With READ_ONCE() the 64 bits are being read:
>>
>> Without the READ_ONCE() only 32 bits are read. That's the most optimal.
>>
>> Without READ_ONCE() but with a barrier() after the read, we should get
>> the same result but GCC (GCC 8.1) does less good:
>>
>> Assuming both part of the 64 bits data will fall into a single
>> cacheline, the second read is in the noise.
> 
> They definitely are in the same cacheline.
> 
>> So agreed to drop this change.
> 
> We could be smart about this and force the compiler to issue a 32bit
> read for 32bit builds. See below. Not sure whether it's worth it, but
> OTOH it will take quite a while until the 32bit time interfaces die
> completely.

I don't think it is worth something so big to just save 1 or 2 cycles in 
time() function. Lets keep it as it is.

Thanks,
Christophe

> 
> Thanks,
> 
>          tglx
> 
> 8<------------
> --- a/include/vdso/datapage.h
> +++ b/include/vdso/datapage.h
> @@ -21,6 +21,18 @@
>   #define CS_RAW		1
>   #define CS_BASES	(CS_RAW + 1)
>   
> +#ifdef __LITTLE_ENDIAN
> +struct sec_hl {
> +	u32	sec_l;
> +	u32	sec_h;
> +};
> +#else
> +struct sec_hl {
> +	u32	sec_h;
> +	u32	sec_l;
> +};
> +#endif
> +
>   /**
>    * struct vdso_timestamp - basetime per clock_id
>    * @sec:	seconds
> @@ -35,7 +47,10 @@
>    * vdso_data.cs[x].shift.
>    */
>   struct vdso_timestamp {
> -	u64	sec;
> +	union {
> +		u64		sec;
> +		struct sec_hl	sec_hl;
> +	};
>   	u64	nsec;
>   };
>   
> --- a/lib/vdso/gettimeofday.c
> +++ b/lib/vdso/gettimeofday.c
> @@ -165,8 +165,13 @@ static __maybe_unused int
>   static __maybe_unused __kernel_old_time_t __cvdso_time(__kernel_old_time_t *time)
>   {
>   	const struct vdso_data *vd = __arch_get_vdso_data();
> -	__kernel_old_time_t t = READ_ONCE(vd[CS_HRES_COARSE].basetime[CLOCK_REALTIME].sec);
> +	__kernel_old_time_t t;
>   
> +#if BITS_PER_LONG == 32
> +	t = READ_ONCE(vd[CS_HRES_COARSE].basetime[CLOCK_REALTIME].sec_hl.sec_l);
> +#else
> +	t = READ_ONCE(vd[CS_HRES_COARSE].basetime[CLOCK_REALTIME].sec);
> +#endif
>   	if (time)
>   		*time = t;
>   
>