[U-Boot] [RFC PATCH] ARM: cache: cp15: Align addresses when initial page_table setup is flushed

[Lukasz Majewski]

Dear all,

> Change made in the commit:
> "arm: Show cache warnings in U-Boot proper only"
> SHA1: bcc53bf095893fbdae531a9a7b5d4ef4a125a7fc
> 
> has revealed that during initial setting of MMU regions in the
> mmu_set_region_dcache_behavior() function some addresses are unaligned
> to platform cache line size.
> 
> As a result we were experiencing following warning messages at early
> boot: CACHE: Misaligned operation at range [8fff0000, 8fff0004]
> CACHE: Misaligned operation at range [8fff0024, 8fff0028]
> 
> Those were caused by an attempt to update single page_table
> (gd->arch.tlb_addr) entries with proper TLB cache settings.
> Since TLB section covers large area (up to 2MiB), we had to update
> very small amount of cache data, very often much smaller than single
> cache line size (e.g. 32 or 64 bytes).
> 
> This patch squashes this warning by properly aligning start and end
> addresses. In fact it does what cache HW would do anyway (flush the
> whole data cache lines).
> Even without this patch it all worked, because TLB table sections were
> initialized to default values earlier.
> 
> Signed-off-by: Lukasz Majewski <l.majewski at majess.pl>
> ---
>  arch/arm/lib/cache-cp15.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/lib/cache-cp15.c b/arch/arm/lib/cache-cp15.c
> index 1121dc3..913c554 100644
> --- a/arch/arm/lib/cache-cp15.c
> +++ b/arch/arm/lib/cache-cp15.c
> @@ -62,6 +62,7 @@ void mmu_set_region_dcache_behaviour(phys_addr_t
> start, size_t size, enum dcache_option option)
>  {
>  	u32 *page_table = (u32 *)gd->arch.tlb_addr;
> +	u32 align_start_addr, align_end_addr;
>  	unsigned long upto, end;
>  
>  	end = ALIGN(start + size, MMU_SECTION_SIZE) >>
> MMU_SECTION_SHIFT; @@ -70,7 +71,14 @@ void
> mmu_set_region_dcache_behaviour(phys_addr_t start, size_t size,
> option); for (upto = start; upto < end; upto++)
>  		set_section_dcache(upto, option);
> -	mmu_page_table_flush((u32)&page_table[start],
> (u32)&page_table[end]); +
> +	align_start_addr = (u32)&page_table[start]
> +		& ~(CONFIG_SYS_CACHELINE_SIZE - 1);
> +	align_end_addr = ALIGN((u32)&page_table[end],
> +			       CONFIG_SYS_CACHELINE_SIZE);

This patch is an RFC on purpose :-).

It requires the CONFIG_SYS_CACHELINE_SIZE to be defined for the target
platform.
However, it happens that some older boards (like Samsung's smdk2410,
Siemens amx) don't have this define, which means that I would need to
define it for each board.

The other option is to get the cache line size from coprocessor (like
get_ccsidr() @ arch/arm/cpu/armv7/cache_v7.c). Such approach would
require some common code extraction.

Best regards,
Łukasz Majewski 


> +	debug("%s: align_start_addr: 0x%x align end_addr: 0x%x\n",
> __func__,
> +	      align_start_addr, align_end_addr);
> +	mmu_page_table_flush(align_start_addr, align_end_addr);
>  }
>  
>  __weak void dram_bank_mmu_setup(int bank)

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: OpenPGP digital signature
URL: <http://lists.denx.de/pipermail/u-boot/attachments/20160809/d3e56764/attachment.sig>