Received-SPF: pass (google.com: domain of linux-kernel+bounces-101449-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:40f1:3f00::1 as permitted sender) client-ip=2604:1380:40f1:3f00::1;
Feedback-ID: i56a14606:Fastmail
User-Agent: Cyrus-JMAP/3.11.0-alpha0-251-g8332da0bf6-fm-20240305.001-g8332da0b
Precedence: bulk
MIME-Version: 1.0
Message-Id: <e0684202-6926-4bd1-86f6-2bb682524712@app.fastmail.com>
In-Reply-To: <4f945f4a-ac29-3ef7-9e15-123962f2a0e9@huawei.com>
References: <20240313084707.3292300-1-arnd@kernel.org>
 <4f945f4a-ac29-3ef7-9e15-123962f2a0e9@huawei.com>
Date: Wed, 13 Mar 2024 12:53:08 +0100
From: "Arnd Bergmann" <arnd@arndb.de>
To: "Zhihao Cheng" <chengzhihao1@huawei.com>,
 "Arnd Bergmann" <arnd@kernel.org>, "Richard Weinberger" <richard@nod.at>,
 "Miquel Raynal" <miquel.raynal@bootlin.com>,
 "Vignesh Raghavendra" <vigneshr@ti.com>,
 "Daniel Golle" <daniel@makrotopia.org>
Cc: linux-mtd@lists.infradead.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] mtd: ubi: avoid expensive do_div() on 32-bit machines
Content-Type: text/plain;charset=utf-8
Content-Transfer-Encoding: quoted-printable

On Wed, Mar 13, 2024, at 12:29, Zhihao Cheng wrote:
> =E5=9C=A8 2024/3/13 16:46, Arnd Bergmann =E5=86=99=E9=81=93:
>> From: Arnd Bergmann <arnd@arndb.de>
>>=20
>> The use of do_div() in ubi_nvmem_reg_read() makes calling it on
>> 32-bit machines rather expensive. Since the 'from' variable is
>> known to be a 32-bit quantity, it is clearly never needed and
>> can be optimized into a regular division operation.
>>=20
> Do you meet a performance problem on a 32-bit machine? There are too=20
> many places invoking do_div, why do you optimize this one?
> Have you tested the influence on a x86_64 platform after this patch=20
> applied? Looks like that do_div is more efficient in x86.

This one was just introduced. The call site looks like a fast
path and it caused a build regression that Daniel addressed with
an suboptimal commit b8a77b9a5f9c ("mtd: ubi: fix NVMEM over
UBI volumes on 32-bit systems").

The way it usually goes is that someone adds an open-coded
64-bit division that causes a link failure, which prompts
the original developer to either rewrite the code to avoid
the long division if possible, or add do_div() after showing
that it is now performance critical, e.g. only called at
probe time.

     Arnd