[PATCH] arm64: mm: check length in sync_icache_aliases for performance

[Mark Rutland]

Hi,

On Thu, May 11, 2017 at 04:19:32PM +0800, Shaokun Zhang wrote:
> sync_icache_aliases calls flush_icache_range if icache is non-aliasing
> policy[see 0a28714 ("arm64: Use PoU cache instr for I/D coherency")].
>   
> If icache uses non-aliasing and page size is 64K, it will broadcast 1K
> DVMs(IC IVAU) to other cpu cores per page. In multi-cores system, so many
> DVMs would degenerate performance. Even if page size is 4K, 64 DVMs will
> be broadcasted and executed.

Please note that this depends on the I-cache and D-cache line sizes,
which are not necessarily 64 bytes.

This is also dependent on system integration. DVMs are not an
architectural concept, and the interconnect may optimize this (e.g. with
snoop filters).

> This patch fixes this issue using invalidation icache all instread of by
> VA when length is one or multiple PAGE_SIZE, especailly for
> __sync_icache_dcache.

This means that we'll over-invalidate the I-caches all the time,
potentially harming the performance of unrelated tasks. So this is not
necessarily an improvement.

Do you have a particular workload which is affected by this?

Thanks,
Mark.

> Signed-off-by: Shaokun Zhang <zhangshaokun at hisilicon.com>
> ---
>  arch/arm64/mm/flush.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/flush.c b/arch/arm64/mm/flush.c
> index 21a8d82..f71da2d 100644
> --- a/arch/arm64/mm/flush.c
> +++ b/arch/arm64/mm/flush.c
> @@ -29,7 +29,7 @@ void sync_icache_aliases(void *kaddr, unsigned long len)
>  {
>  	unsigned long addr = (unsigned long)kaddr;
>  
> -	if (icache_is_aliasing()) {
> +	if ((len >= PAGE_SIZE) || icache_is_aliasing()) {
>  		__clean_dcache_area_pou(kaddr, len);
>  		__flush_icache_all();
>  	} else {
> -- 
> 1.9.1
>