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] powerpc/32: Switch VDSO to C implementation.
To:     Thomas Gleixner <tglx@linutronix.de>
Cc:     Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Paul Mackerras <paulus@samba.org>,
        Michael Ellerman <mpe@ellerman.id.au>,
        vincenzo.frascino@arm.com, luto@kernel.org,
        linux-kernel@vger.kernel.org, linuxppc-dev@lists.ozlabs.org
References: <8ce3582f7f7da9ff0286ced857e5aa2e5ae6746e.1571662378.git.christophe.leroy@c-s.fr>
 <alpine.DEB.2.21.1910212312520.2078@nanos.tec.linutronix.de>
 <f4486e86-3c0c-0eec-1639-0e5956cdb8f1@c-s.fr>
From:   Christophe Leroy <christophe.leroy@c-s.fr>
Message-ID: <95bd2367-8edc-29db-faa3-7729661e05f2@c-s.fr>
Date:   Tue, 22 Oct 2019 15:56:33 +0200
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:60.0) Gecko/20100101
 Thunderbird/60.9.0
MIME-Version: 1.0
In-Reply-To: <f4486e86-3c0c-0eec-1639-0e5956cdb8f1@c-s.fr>
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 22/10/2019 à 11:01, Christophe Leroy a écrit :
> 
> 
> Le 21/10/2019 à 23:29, Thomas Gleixner a écrit :
>> On Mon, 21 Oct 2019, Christophe Leroy wrote:
>>
>>> This is a tentative to switch powerpc/32 vdso to generic C 
>>> implementation.
>>> It will likely not work on 64 bits or even build properly at the moment.
>>>
>>> powerpc is a bit special for VDSO as well as system calls in the
>>> way that it requires setting CR SO bit which cannot be done in C.
>>> Therefore, entry/exit and fallback needs to be performed in ASM.
>>>
>>> To allow that, C fallbacks just return -1 and the ASM entry point
>>> performs the system call when the C function returns -1.
>>>
>>> The performance is rather disappoiting. That's most likely all
>>> calculation in the C implementation are based on 64 bits math and
>>> converted to 32 bits at the very end. I guess C implementation should
>>> use 32 bits math like the assembly VDSO does as of today.
>>
>>> gettimeofday:    vdso: 750 nsec/call
>>>
>>> gettimeofday:    vdso: 1533 nsec/call
> 
> Small improvement (3%) with the proposed change:
> 
> gettimeofday:    vdso: 1485 nsec/call

By inlining do_hres() I get the following:

gettimeofday:    vdso: 1072 nsec/call

Christophe

> 
> Though still some way to go.
> 
> Christophe
> 
>>
>> The only real 64bit math which can matter is the 64bit * 32bit multiply,
>> i.e.
>>
>> static __always_inline
>> u64 vdso_calc_delta(u64 cycles, u64 last, u64 mask, u32 mult)
>> {
>>          return ((cycles - last) & mask) * mult;
>> }
>>
>> Everything else is trivial add/sub/shift, which should be roughly the 
>> same
>> in ASM.
>>
>> Can you try to replace that with:
>>
>> static __always_inline
>> u64 vdso_calc_delta(u64 cycles, u64 last, u64 mask, u32 mult)
>> {
>>          u64 ret, delta = ((cycles - last) & mask);
>>          u32 dh, dl;
>>
>>          dl = delta;
>>          dh = delta >> 32;
>>
>>          res = mul_u32_u32(al, mul);
>>          if (ah)
>>                  res += mul_u32_u32(ah, mul) << 32;
>>
>>          return res;
>> }
>>
>> That's pretty much what __do_get_tspec does in ASM.
>>
>> Thanks,
>>
>>     tglx
>>