On Mon, Nov 27, 2017 at 1:27 PM, Greentime Hu <green.hu@xxxxxxxxx> wrote:
> From: Greentime Hu <greentime@xxxxxxxxxxxxx>
>
> This patch introduces ioremap implementations.
>
> Signed-off-by: Vincent Chen <vincentc@xxxxxxxxxxxxx>
> Signed-off-by: Greentime Hu <greentime@xxxxxxxxxxxxx>
> ---
> arch/nds32/include/asm/io.h | 25 +++++++++++++++
> arch/nds32/mm/ioremap.c | 75 +++++++++++++++++++++++++++++++++++++++++++
> 2 files changed, 100 insertions(+)
> create mode 100644 arch/nds32/include/asm/io.h
> create mode 100644 arch/nds32/mm/ioremap.c
>
> diff --git a/arch/nds32/include/asm/io.h b/arch/nds32/include/asm/io.h
> new file mode 100644
> index 0000000..b83dea1
> --- /dev/null
> +++ b/arch/nds32/include/asm/io.h
> @@ -0,0 +1,25 @@
> +#ifndef __ASM_NDS32_IO_H
> +#define __ASM_NDS32_IO_H
> +
> +#ifdef __KERNEL__
> +void iounmap(void __iomem * addr);
> +#include <asm-generic/io.h>
> +
> +#endif /* __KERNEL__ */
> +#endif /* __ASM_NDS32_IO_H */
Here, you should define a lot of the I/O accessor functions,
dereferencing a pointer
is generally not enough to guarantee an atomic MMIO operation. You need to
force the access to use the correct size to prevent the compiler from issuing
byte-sized operations when it thinks the pointer might be unaligned, and
there should be barriers that ensure a memory access is synchronized with
a DMA that might be triggered by a writel, or claimed to be completed after
a readl. Please see the risc-v header for this, it has many good explanations.
Arnd
