Feedback-ID: i56a14606:Fastmail
User-Agent: Cyrus-JMAP/3.9.0-alpha0-172-g9a2dae1853-fm-20230213.001-g9a2dae18
Mime-Version: 1.0
Message-Id: <f3e42012-609c-4085-b4f4-bd32bfc34aff@app.fastmail.com>
In-Reply-To: <20230227142809.kujmrraf3pcdhqyn@CAB-WSD-L081021>
References: <20230222115020.55867-1-avromanov@sberdevices.ru>
 <8e5f9bfa-d612-cd43-d722-d04c40938c62@linaro.org>
 <20230227142809.kujmrraf3pcdhqyn@CAB-WSD-L081021>
Date:   Mon, 27 Feb 2023 15:58:50 +0100
From:   "Arnd Bergmann" <arnd@arndb.de>
To:     "Dmitry Rokosov" <ddrokosov@sberdevices.ru>,
        "Neil Armstrong" <neil.armstrong@linaro.org>
Cc:     "Alexey Romanov" <avromanov@sberdevices.ru>,
        "Rob Herring" <robh+dt@kernel.org>,
        krzysztof.kozlowski+dt@linaro.org,
        "Kevin Hilman" <khilman@baylibre.com>, jbrunet@baylibre.com,
        martin.blumenstingl@googlemail.com,
        "Linus Walleij" <linus.walleij@linaro.org>,
        devicetree@vger.kernel.org, linux-arm-kernel@lists.infradead.org,
        linux-amlogic@lists.infradead.org, linux-kernel@vger.kernel.org,
        kernel@sberdevices.ru
Subject: Re: [PATCH v1 0/3] Meson A1 32-bit support
Content-Type: text/plain
Precedence: bulk

On Mon, Feb 27, 2023, at 15:28, Dmitry Rokosov wrote:
> Hello Neil!
>
> On Mon, Feb 27, 2023 at 09:15:04AM +0100, neil.armstrong@linaro.org wrote:
>
> [...]
>
>> I'm aware Amlogic also runs their kernel as 32bit to gain a few kbytes
>> of memory, but those processors are ARMv8 and the arm64 arch code
>> has been designed for those CPUs.
>> 
>> So far I didn't find a single good reason to add 32bit support for
>> ARMv8 Amlogic based SoCs, if you have a solid reason please share.
>
> I totally agree with you, but I suppose it's fully related to 'big'
> Amlogic SoC like S905_ or A311_ series. A113L (aka 'a1') is
> a cost-efficient dual-core SoC which is used for small, cheap solutions
> with cheap components. Every cent is important during BoM development.
> That's why usually ODMs install small ROM and RAM capacity, and each
> megabyte is important for RAM/ROM kernel and rootfs footprints.
> Why am I talking about rootfs? For such small projects a good
> choice is buildroot rootfs assembling framework. Unfortunatelly,
> buildroot doesn't support 'compat' mode when kernel and userspace have
> a different bitness. In the internal project, we save several
> percents of ROM/RAM free space using 32-bit configuration (mostly rootfs
> ROM space, to be honest). Therefore, for such 'little' cost-efficient
> SoCs we can make an exception and support 32-bit configuration, from my
> point of view.
>
> What do you think about that?

I would argue that is a problem with buildroot, and using a 32-bit
kernel is not something we should encourage over fixing buildroot
to do it right, or building the kernel separately from the rootfs.

We do allow building support for a couple of ARMv8 SoCs in 32-bit
mode, but that is usually because they ship with a 32-bit bootrom
and cannot actually run a 64-bit kernel.

The overhead of running a 64-bit kernel is usually a few megabytes
compared to a 32-bit kernel, to store the larger kernel .text/.data
segments, per-thread stack and page tables as well as 'page',
and 'inode' structures. I see that A1 only supports DDR3 and DDR4
memory, so I assume that there are always at least 128MB
of total RAM available, or 512MB for the most cost-effective
size with a single memory chip.

My feeling is that for the 256MB configuration, it is very hard to
argue for a 32-bit kernel because of the countless downsides,
and even for the 128MB configuration, I would still try to avoid
it out of principle.

      Arnd