Re: [2.5 PATCH] R10K DMA cache flushing routines for non-coherent systems

[Ralf Baechle <ralf@oss.sgi.com>]

On Sun, Jul 14, 2002 at 11:17:39PM +0200, Vivien Chappelier wrote:

> @@ -111,13 +115,96 @@
>       }
>  }
>  
> +static void
> +andes_flush_cache_all(void)
> +{
> +     andes_flush_cache_l1();
> +     andes_flush_cache_l2();
> +}

We can optimize that slightly.  By leaving it an empty function as it is :-)

>  void
>  andes_flush_icache_page(unsigned long page)
>  {
> -     if (scache_lsz64)
> -             blast_scache64_page(page);
> -     else
> -             blast_scache128_page(page);
> +     switch (sc_lsize) {
> +             case 64:
> +                     blast_scache64_page(page);
> +                     break;
> +             case 128:
> +                     blast_scache128_page(page);
> +                     break;

Eh...

So this is replacing a wrong version with another wrong piece of code.
There simply is no reason to flush the second level cache at this point.
That forcing instructions to be re-fetched from memory and that's slooow.

> +     _dma_cache_wback_inv = andes_dma_cache_wback_inv;
> +     _dma_cache_wback = andes_dma_cache_wback;
> +     _dma_cache_inv = andes_dma_cache_inv;

This is breaking cache coherent machines.

Encore une fois :-)

  Ralf