Re: [PATCH v2] staging: trivial: hikey9xx: fix be32<->u32 casting warnings

[Arnd Bergmann <arnd@kernel.org>]

On Thu, Nov 19, 2020 at 1:31 PM Juan Antonio Aldea-Armenteros
<juant.aldea@gmail.com> wrote:
>
> This patch fixes the following warnings reported by sparse, by adding
> missing __force annotations.
>
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
> drivers/staging/hikey9xx/hisi-spmi-controller.c:164:24: warning: cast to restricted __be32
>
> drivers/staging/hikey9xx/hisi-spmi-controller.c:239:25: warning: cast from restricted __be32
>
> Rationale for #164:
> data is declared as u32, and it is read and then converted by means of
> be32_to_cpu(). Said function expects a __be32 but data is u32, therefore
> there's a type missmatch here.
>
> Rationale for #239:
> Is the dual of #164. This time data going to be  written so it
> needs to be converted from cpu to __be32, but writel() expects u32 and the
> output of cpu_to_be32 returns a __be32.

Both of the casts look very suspicious, I'd leave these in unless
someone can confirm what the actual desired behavior is.

>                              SPMI_SLAVE_OFFSET * slave_id +
>                              SPMI_APB_SPMI_RDATA0_BASE_ADDR +
>                              i * SPMI_PER_DATAREG_BYTE);
> -               data = be32_to_cpu((__be32)data);
> +               data = be32_to_cpu((__be32 __force)data);
>                 if ((bc - i * SPMI_PER_DATAREG_BYTE) >> 2) {
>                         memcpy(buf, &data, sizeof(data));
>                         buf += sizeof(data);

The data comes from a readl(), which contains an endian conversion
on architectures that need it, such as when running the board
in big-endian arm64 mode. Having a second endian-conversion
on little-endian architectures means that the data is always swapped
when it gets written to the register.

In the original code before Mauro's commit 8788a30c12c7 ("staging:
spmi: hisi-spmi-controller: use le32 macros where needed"), the data
was byteswapped, and then written into the fifo register, which
produced no warning but would do a double-swap on a big-endian
kernel, and change the behavior from what it was before.

My guess is that Mauro inadvertently fixed this driver for big-endian
mode, without noticing that it was broken to start with, and that
he did not actually try it with CONFIG_CPU_BIG_ENDIAN.

I think the best way would be to go back to to using swab32p()
(not the open-coded version) and then use writesl() or
iowrite32_rep() with count=1 to write the byteswapped FIFO
register without swapping it again.

      Arnd