Re: [PATCH v4 2/6] x86: document WC MTRR effects on PAT / non-PAT pages

[Randy Dunlap <rdunlap@infradead.org>]

On 04/29/15 14:44, Luis R. Rodriguez wrote:
> From: "Luis R. Rodriguez" <mcgrof@suse.com>
> 

> ---
>  Documentation/x86/mtrr.txt      | 18 +++++++++++++++---
>  Documentation/x86/pat.txt       | 40 +++++++++++++++++++++++++++++++++++++++-
>  arch/x86/kernel/cpu/mtrr/main.c |  3 +++
>  3 files changed, 57 insertions(+), 4 deletions(-)
> 
> diff --git a/Documentation/x86/pat.txt b/Documentation/x86/pat.txt
> index cf08c9f..7e183e3 100644
> --- a/Documentation/x86/pat.txt
> +++ b/Documentation/x86/pat.txt
> @@ -102,7 +104,43 @@ wants to export a RAM region, it has to do set_memory_uc() or set_memory_wc()
>  as step 0 above and also track the usage of those pages and use set_memory_wb()
>  before the page is freed to free pool.
>  
> -
> +MTRR effects on PAT / non-PAT systems
> +-------------------------------------
> +
> +The following table provides the effects of using write-combining MTRRs when
> +using ioremap*() calls on x86 for both non-PAT and PAT systems. Ideally
> +mtrr_add() usage will be phased in favor of arch_phys_wc_add() which will
> +be a no-op on PAT enabled systems. The region over which a arch_phys_wc_add()
> +is made should already have be ioremap'd with write-combining page attributes
> +or PAT entries, this can be done by using ioremap_wc() / or respective helpers.
> +Devices which combine areas of IO memory desired to remain uncachable with

I would spell it uncacheable.  In kernel Documentation/, grep uncacheable finds
14 hits vs. 6 hits for uncachable.  No big deal.

> +areas where write-combining is desirable and are restricted by the size
> +requirements of MTRRs should consider splitting up their IO memory space
> +cleanly with ioremap_uc() and ioremap_wc() followed by an arch_phys_wc_add()
> +encompassing both regions. Such use is nevertheless heavily discouraged as
> +the effective memory type is considered implementation defined. This strategy
> +should only be used as last resort on devices with size-contrained regions

                                                      size-constrained

> +where otherwise MTRR write-combining would not be effective.
> +
> +Note that you cannot use set_memory_wc() to override / whitelist IO remapped
> +memory space mapped with ioremap*() calls, set_memory_wc() can only be used
> +on RAM.
> +
> +----------------------------------------------------------------------
> +MTRR Non-PAT   PAT    Linux ioremap value        Effective memory type
> +----------------------------------------------------------------------
> +                                                  Non-PAT |  PAT
> +     PAT
> +     |PCD
> +     ||PWT
> +     |||
> +WC   000      WB      _PAGE_CACHE_MODE_WB            WC   |   WC
> +WC   001      WC      _PAGE_CACHE_MODE_WC            WC*  |   WC
> +WC   010      UC-     _PAGE_CACHE_MODE_UC_MINUS      WC*  |   WC
> +WC   011      UC      _PAGE_CACHE_MODE_UC            UC   |   UC
> +----------------------------------------------------------------------
> +
> +(*) denotes implementation defined and is discouraged
>  
>  Notes:
>  
> diff --git a/arch/x86/kernel/cpu/mtrr/main.c b/arch/x86/kernel/cpu/mtrr/main.c
> index ea5f363..12abdbe 100644
> --- a/arch/x86/kernel/cpu/mtrr/main.c
> +++ b/arch/x86/kernel/cpu/mtrr/main.c
> @@ -538,6 +538,9 @@ EXPORT_SYMBOL(mtrr_del);
>   * attempts to add a WC MTRR covering size bytes starting at base and
>   * logs an error if this fails.
>   *
> + * The caller should expect to need to provide a power of two size on an

    * The called should provide a power of two size on an equivalent
    * power of two boundary.

> + * equivalent power of two boundary.
> + *
>   * Drivers must store the return value to pass to mtrr_del_wc_if_needed,
>   * but drivers should not try to interpret that return value.
>   */
> 


-- 
~Randy