Re: [RFT PATCH v3 08/27] asm-generic/io.h: Add a non-posted variant of ioremap()

[Will Deacon]

On Fri, Mar 05, 2021 at 06:38:43AM +0900, Hector Martin wrote:
> ARM64 currently defaults to posted MMIO (nGnRnE), but some devices
> require the use of non-posted MMIO (nGnRE). Introduce a new ioremap()
> variant to handle this case. ioremap_np() is aliased to ioremap() by
> default on arches that do not implement this variant.
> 
> sparc64 is the only architecture that needs to be touched directly,
> because it includes neither of the generic io.h or iomap.h headers.
> 
> This adds the IORESOURCE_MEM_NONPOSTED flag, which maps to this
> variant and marks a given resource as requiring non-posted mappings.
> This is implemented in the resource system because it is a SoC-level
> requirement, so existing drivers do not need special-case code to pick
> this ioremap variant.
> 
> Then this is implemented in devres by introducing devm_ioremap_np(),
> and making devm_ioremap_resource() automatically select this variant
> when the resource has the IORESOURCE_MEM_NONPOSTED flag set.
> 
> Signed-off-by: Hector Martin <marcan@xxxxxxxxx>
> ---
>  .../driver-api/driver-model/devres.rst        |  1 +
>  arch/sparc/include/asm/io_64.h                |  4 ++++
>  include/asm-generic/io.h                      | 22 ++++++++++++++++++-
>  include/asm-generic/iomap.h                   |  9 ++++++++
>  include/linux/io.h                            |  2 ++
>  include/linux/ioport.h                        |  1 +
>  lib/devres.c                                  | 22 +++++++++++++++++++
>  7 files changed, 60 insertions(+), 1 deletion(-)
> 
> diff --git a/Documentation/driver-api/driver-model/devres.rst b/Documentation/driver-api/driver-model/devres.rst
> index cd8b6e657b94..2f45877a539d 100644
> --- a/Documentation/driver-api/driver-model/devres.rst
> +++ b/Documentation/driver-api/driver-model/devres.rst
> @@ -309,6 +309,7 @@ IOMAP
>    devm_ioremap()
>    devm_ioremap_uc()
>    devm_ioremap_wc()
> +  devm_ioremap_np()
>    devm_ioremap_resource() : checks resource, requests memory region, ioremaps
>    devm_ioremap_resource_wc()
>    devm_platform_ioremap_resource() : calls devm_ioremap_resource() for platform device
> diff --git a/arch/sparc/include/asm/io_64.h b/arch/sparc/include/asm/io_64.h
> index 9bb27e5c22f1..9fbfc9574432 100644
> --- a/arch/sparc/include/asm/io_64.h
> +++ b/arch/sparc/include/asm/io_64.h
> @@ -409,6 +409,10 @@ static inline void __iomem *ioremap(unsigned long offset, unsigned long size)
>  #define ioremap_uc(X,Y)			ioremap((X),(Y))
>  #define ioremap_wc(X,Y)			ioremap((X),(Y))
>  #define ioremap_wt(X,Y)			ioremap((X),(Y))
> +static inline void __iomem *ioremap_np(unsigned long offset, unsigned long size)
> +{
> +	return NULL;
> +}
>  
>  static inline void iounmap(volatile void __iomem *addr)
>  {
> diff --git a/include/asm-generic/io.h b/include/asm-generic/io.h
> index c6af40ce03be..082e0c96db6e 100644
> --- a/include/asm-generic/io.h
> +++ b/include/asm-generic/io.h
> @@ -942,7 +942,9 @@ static inline void *phys_to_virt(unsigned long address)
>   *
>   * ioremap_wc() and ioremap_wt() can provide more relaxed caching attributes
>   * for specific drivers if the architecture choses to implement them.  If they
> - * are not implemented we fall back to plain ioremap.
> + * are not implemented we fall back to plain ioremap. Conversely, ioremap_np()
> + * can provide stricter non-posted write semantics if the architecture
> + * implements them.
>   */
>  #ifndef CONFIG_MMU
>  #ifndef ioremap
> @@ -993,6 +995,24 @@ static inline void __iomem *ioremap_uc(phys_addr_t offset, size_t size)
>  {
>  	return NULL;
>  }
> +
> +/*
> + * ioremap_np needs an explicit architecture implementation, as it
> + * requests stronger semantics than regular ioremap(). Portable drivers
> + * should instead use one of the higher-level abstractions, like
> + * devm_ioremap_resource(), to choose the correct variant for any given
> + * device and bus. Portable drivers with a good reason to want non-posted
> + * write semantics should always provide an ioremap() fallback in case
> + * ioremap_np() is not available.
> + */
> +#ifndef ioremap_np
> +#define ioremap_np ioremap_np
> +static inline void __iomem *ioremap_np(phys_addr_t offset, size_t size)
> +{
> +	return NULL;
> +}
> +#endif

Can we implement the generic pci_remap_cfgspace() in terms of ioremap_np()
if it is supported by the architecture? That way, we could avoid defining
both on arm64.

Will