Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Feedback-ID: i56a14606:Fastmail
User-Agent: Cyrus-JMAP/3.7.0-alpha0-1185-g841157300a-fm-20221208.002-g84115730
Mime-Version: 1.0
Message-Id: <9105d6fc-880b-4734-857d-e3d30b87ccf6@app.fastmail.com>
In-Reply-To: <20230109104054.gabdvjxyawdfbqii@skbuf>
References: <20230109104054.gabdvjxyawdfbqii@skbuf>
Date:   Mon, 09 Jan 2023 12:33:31 +0100
From:   "Arnd Bergmann" <arnd@arndb.de>
To:     "Vladimir Oltean" <vladimir.oltean@nxp.com>
Cc:     Linux-Arch <linux-arch@vger.kernel.org>,
        linux-kernel@vger.kernel.org
Subject: Re: Sparse warning when using ioread64() from include/asm-generic/io.h
Content-Type: text/plain
Precedence: bulk

On Mon, Jan 9, 2023, at 11:40, Vladimir Oltean wrote:
> Hi,
>
> I would like to get rid of the following sparse error in the enetc
> driver (for arm64), which uses ioread64().
>
> ../drivers/net/ethernet/freescale/enetc/enetc_ethtool.c: note: in included file
> 	(through ../arch/arm64/include/asm/io.h, ../include/linux/io.h,
> 	../include/linux/irq.h, ../include/asm-generic/hardirq.h,
> 	../arch/arm64/include/asm/hardirq.h, ...):
> ../include/asm-generic/io.h:239:15: warning: cast to restricted __le64
>
> The trouble is I don't understand why the casts to __le64 and use of
> __le64_to_cpu() are even needed, when everything seems to be native
> endianness. I've seen commit c1d55d50139b ("asm-generic/io.h: Fix sparse
> warnings on big-endian architectures"), but that doesn't claim to fix
> anything for little endian (and doesn't touch the 64 accessors, for some
> reason).
>
> Could you please help?

From what I can tell, the fix for openrisc was described as
a big-endian warning fix, but the warning is actually the same
on both. The difference is that on little-endian kernels,
the __le64_to_cpu() conversion only changes the type but not
the value, while on big-endian machines, the value would
be wrong without the conversion: __raw_readl() is defined
to never byteswap the data, while readl() must byteswap
little-endian MMIO registers into big-endian CPU registers
when CONFIG_CPU_BIG_ENDIAN is set.

Since Stafford only tested on 32-bit OpenRISC, he missed the
readq()/writeq() accessors that need the same warning fix.

     Arnd