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 12/13] powerpc/kernel: Do not inconditionally save
 non volatile registers on system call
To:     Nicholas Piggin <npiggin@gmail.com>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Michael Ellerman <mpe@ellerman.id.au>, msuchanek@suse.de,
        Paul Mackerras <paulus@samba.org>
Cc:     linux-kernel@vger.kernel.org, linuxppc-dev@lists.ozlabs.org
References: <029e1064b1ad738785718221ea468c9cfc282457.1586108649.git.christophe.leroy@c-s.fr>
 <4ef6d617cfd34e09e9bf5a456b2e0b6d2a8a3c96.1586108649.git.christophe.leroy@c-s.fr>
 <1586135554.pnqaj0giue.astroid@bobo.none>
From:   Christophe Leroy <christophe.leroy@c-s.fr>
Message-ID: <48d3e2e5-318f-011e-a59b-ec89bd7b76d2@c-s.fr>
Date:   Mon, 6 Apr 2020 20:18:29 +0200
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.6.0
MIME-Version: 1.0
In-Reply-To: <1586135554.pnqaj0giue.astroid@bobo.none>
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 06/04/2020 à 03:25, Nicholas Piggin a écrit :
> Christophe Leroy's on April 6, 2020 3:44 am:
>> Before : 347 cycles on null_syscall
>> After  : 327 cycles on null_syscall
> 
> The problem I had doing this is that signal delivery wnats full regs,
> and you don't know if you have a signal pending ahead of time if you
> have interrupts enabled.
> 
> I began to try bailing out back to asm to save nvgprs and call again.
> I think that can be made to work, but it is more complication in asm,
> and I soon found that 64s CPUs don't care about NVGPRs too much so it's
> nice to get rid of the !fullregs state.

I tried a new way in v3, please have a look. I split 
syscall_exit_prepare() in 3 parts and the result is unexpected: it is 
better than before the series (307 cycles now versus 311 cycles with 
full ASM syscall entry/exit).

> 
> Possibly another approach would be to leave interrupts disabled for the
> case where you have no work to do. You could create a small
> syscall_exit_prepare_nowork fastpath for that case for 32-bit, perhaps?
> 
> Thanks,
> Nick
> 

Christophe